xref: /titanic_51/usr/src/uts/common/io/cxgbe/t4nex/t4_sge.c (revision 56b2bdd1f04d465cfe4a95b88ae5cba5884154e4)
1 /*
2  * This file and its contents are supplied under the terms of the
3  * Common Development and Distribution License ("CDDL"), version 1.0.
4  * You may only use this file in accordance with the terms of version
5  * 1.0 of the CDDL.
6  *
7  * A full copy of the text of the CDDL should have accompanied this
8  * source. A copy of the CDDL is also available via the Internet at
9  * http://www.illumos.org/license/CDDL.
10  */
11 
12 /*
13  * This file is part of the Chelsio T4 support code.
14  *
15  * Copyright (C) 2010-2013 Chelsio Communications.  All rights reserved.
16  *
17  * This program is distributed in the hope that it will be useful, but WITHOUT
18  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
19  * FITNESS FOR A PARTICULAR PURPOSE.  See the LICENSE file included in this
20  * release for licensing terms and conditions.
21  */
22 
23 #include <sys/ddi.h>
24 #include <sys/sunddi.h>
25 #include <sys/sunndi.h>
26 #include <sys/atomic.h>
27 #include <sys/dlpi.h>
28 #include <sys/pattr.h>
29 #include <sys/strsubr.h>
30 #include <sys/stream.h>
31 #include <sys/strsun.h>
32 #include <inet/ip.h>
33 #include <inet/tcp.h>
34 
35 #include "version.h"
36 #include "common/common.h"
37 #include "common/t4_msg.h"
38 #include "common/t4_regs.h"
39 #include "common/t4_regs_values.h"
40 
41 /* TODO: Tune. */
42 int rx_buf_size = 8192;
43 int tx_copy_threshold = 256;
44 uint16_t rx_copy_threshold = 256;
45 
46 enum {
47 	SPG_SIZE = 64,		/* Size of status page */
48 	FL_ALIGN = CACHE_LINE,	/* packet buffer alignment in FL buffers */
49 	FL_PKTSHIFT = 2		/* payload is at this offset in packet buffer */
50 };
51 
52 /* Used to track coalesced tx work request */
53 struct txpkts {
54 	mblk_t *tail;		/* head is in the software descriptor */
55 	uint64_t *flitp;	/* ptr to flit where next pkt should start */
56 	uint8_t npkt;		/* # of packets in this work request */
57 	uint8_t nflits;		/* # of flits used by this work request */
58 	uint16_t plen;		/* total payload (sum of all packets) */
59 };
60 
61 /* All information needed to tx a frame */
62 struct txinfo {
63 	uint32_t len;		/* Total length of frame */
64 	uint32_t flags;		/* Checksum and LSO flags */
65 	uint32_t mss;		/* MSS for LSO */
66 	uint8_t nsegs;		/* # of segments in the SGL, 0 means imm. tx */
67 	uint8_t nflits;		/* # of flits needed for the SGL */
68 	uint8_t hdls_used;	/* # of DMA handles used */
69 	uint32_t txb_used;	/* txb_space used */
70 	struct ulptx_sgl sgl __attribute__((aligned(8)));
71 	struct ulptx_sge_pair reserved[TX_SGL_SEGS / 2];
72 };
73 
74 static int service_iq(struct sge_iq *iq, int budget);
75 static inline void init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx,
76     int8_t pktc_idx, int qsize, uint8_t esize);
77 static inline void init_fl(struct sge_fl *fl, uint16_t qsize);
78 static inline void init_eq(struct sge_eq *eq, uint16_t eqtype, uint16_t qsize,
79     uint8_t tx_chan, uint16_t iqid);
80 static int alloc_iq_fl(struct port_info *pi, struct sge_iq *iq,
81     struct sge_fl *fl, int intr_idx, int cong);
82 static int free_iq_fl(struct port_info *pi, struct sge_iq *iq,
83     struct sge_fl *fl);
84 static int alloc_fwq(struct adapter *sc);
85 static int free_fwq(struct adapter *sc);
86 static int alloc_mgmtq(struct adapter *sc);
87 static int alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx,
88     int i);
89 static int free_rxq(struct port_info *pi, struct sge_rxq *rxq);
90 #ifndef TCP_OFFLOAD_DISABLE
91 static int alloc_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq,
92 	int intr_idx);
93 static int free_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq);
94 #endif
95 static int ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq);
96 static int eth_eq_alloc(struct adapter *sc, struct port_info *pi,
97     struct sge_eq *eq);
98 #ifndef TCP_OFFLOAD_DISABLE
99 static int ofld_eq_alloc(struct adapter *sc, struct port_info *pi,
100     struct sge_eq *eq);
101 #endif
102 static int alloc_eq(struct adapter *sc, struct port_info *pi,
103     struct sge_eq *eq);
104 static int free_eq(struct adapter *sc, struct sge_eq *eq);
105 static int alloc_wrq(struct adapter *sc, struct port_info *pi,
106     struct sge_wrq *wrq, int idx);
107 static int free_wrq(struct adapter *sc, struct sge_wrq *wrq);
108 static int alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx);
109 static int free_txq(struct port_info *pi, struct sge_txq *txq);
110 static int alloc_dma_memory(struct adapter *sc, size_t len, int flags,
111     ddi_device_acc_attr_t *acc_attr, ddi_dma_attr_t *dma_attr,
112     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
113     caddr_t *pva);
114 static int free_dma_memory(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl);
115 static int alloc_desc_ring(struct adapter *sc, size_t len, int rw,
116     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
117     caddr_t *pva);
118 static int free_desc_ring(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl);
119 static int alloc_tx_copybuffer(struct adapter *sc, size_t len,
120     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
121     caddr_t *pva);
122 static inline bool is_new_response(const struct sge_iq *iq,
123     struct rsp_ctrl **ctrl);
124 static inline void iq_next(struct sge_iq *iq);
125 static int refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs);
126 static void refill_sfl(void *arg);
127 static void add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl);
128 static void free_fl_bufs(struct sge_fl *fl);
129 static mblk_t *get_fl_payload(struct sge_fl *fl, uint32_t len_newbuf,
130     int *fl_bufs_used);
131 static int get_frame_txinfo(struct sge_txq *txq, mblk_t **fp,
132     struct txinfo *txinfo, int sgl_only);
133 static inline int fits_in_txb(struct sge_txq *txq, int len, int *waste);
134 static inline int copy_into_txb(struct sge_txq *txq, mblk_t *m, int len,
135     struct txinfo *txinfo);
136 static inline void add_seg(struct txinfo *txinfo, uint64_t ba, uint32_t len);
137 static inline int add_mblk(struct sge_txq *txq, struct txinfo *txinfo,
138     mblk_t *m, int len);
139 static void free_txinfo_resources(struct sge_txq *txq, struct txinfo *txinfo);
140 static int add_to_txpkts(struct sge_txq *txq, struct txpkts *txpkts, mblk_t *m,
141     struct txinfo *txinfo);
142 static void write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts);
143 static int write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, mblk_t *m,
144     struct txinfo *txinfo);
145 static inline void write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
146     struct txpkts *txpkts, struct txinfo *txinfo);
147 static inline void copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to,
148     int len);
149 static inline void ring_tx_db(struct adapter *sc, struct sge_eq *eq);
150 static int reclaim_tx_descs(struct sge_txq *txq, int howmany);
151 static void write_txqflush_wr(struct sge_txq *txq);
152 static int t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss,
153     mblk_t *m);
154 static inline void ring_fl_db(struct adapter *sc, struct sge_fl *fl);
155 static kstat_t *setup_port_config_kstats(struct port_info *pi);
156 static kstat_t *setup_port_info_kstats(struct port_info *pi);
157 static kstat_t *setup_rxq_kstats(struct port_info *pi, struct sge_rxq *rxq,
158     int idx);
159 static int update_rxq_kstats(kstat_t *ksp, int rw);
160 static int update_port_info_kstats(kstat_t *ksp, int rw);
161 static kstat_t *setup_txq_kstats(struct port_info *pi, struct sge_txq *txq,
162     int idx);
163 static int update_txq_kstats(kstat_t *ksp, int rw);
164 static int handle_fw_rpl(struct sge_iq *iq, const struct rss_header *rss,
165     mblk_t *m);
166 
167 static inline int
168 reclaimable(struct sge_eq *eq)
169 {
170 	unsigned int cidx;
171 
172 	cidx = eq->spg->cidx;   /* stable snapshot */
173 	cidx = be16_to_cpu(cidx);
174 
175 	if (cidx >= eq->cidx)
176 		return (cidx - eq->cidx);
177 	else
178 		return (cidx + eq->cap - eq->cidx);
179 }
180 
181 void
182 t4_sge_init(struct adapter *sc)
183 {
184 	struct driver_properties *p = &sc->props;
185 	ddi_dma_attr_t *dma_attr;
186 	ddi_device_acc_attr_t *acc_attr;
187 	uint32_t v;
188 
189 	/*
190 	 * Device access and DMA attributes for descriptor rings
191 	 */
192 	acc_attr = &sc->sge.acc_attr_desc;
193 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
194 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
195 	acc_attr->devacc_attr_dataorder = DDI_STRICTORDER_ACC;
196 
197 	dma_attr = &sc->sge.dma_attr_desc;
198 	dma_attr->dma_attr_version = DMA_ATTR_V0;
199 	dma_attr->dma_attr_addr_lo = 0;
200 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
201 	dma_attr->dma_attr_count_max = UINT64_MAX;
202 	dma_attr->dma_attr_align = 512;
203 	dma_attr->dma_attr_burstsizes = 0xfff;
204 	dma_attr->dma_attr_minxfer = 1;
205 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
206 	dma_attr->dma_attr_seg = UINT64_MAX;
207 	dma_attr->dma_attr_sgllen = 1;
208 	dma_attr->dma_attr_granular = 1;
209 	dma_attr->dma_attr_flags = 0;
210 
211 	/*
212 	 * Device access and DMA attributes for tx buffers
213 	 */
214 	acc_attr = &sc->sge.acc_attr_tx;
215 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
216 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
217 
218 	dma_attr = &sc->sge.dma_attr_tx;
219 	dma_attr->dma_attr_version = DMA_ATTR_V0;
220 	dma_attr->dma_attr_addr_lo = 0;
221 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
222 	dma_attr->dma_attr_count_max = UINT64_MAX;
223 	dma_attr->dma_attr_align = 1;
224 	dma_attr->dma_attr_burstsizes = 0xfff;
225 	dma_attr->dma_attr_minxfer = 1;
226 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
227 	dma_attr->dma_attr_seg = UINT64_MAX;
228 	dma_attr->dma_attr_sgllen = TX_SGL_SEGS;
229 	dma_attr->dma_attr_granular = 1;
230 	dma_attr->dma_attr_flags = 0;
231 
232 	/*
233 	 * Device access and DMA attributes for rx buffers
234 	 */
235 	sc->sge.rxb_params.dip = sc->dip;
236 	sc->sge.rxb_params.buf_size = rx_buf_size;
237 
238 	acc_attr = &sc->sge.rxb_params.acc_attr_rx;
239 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
240 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
241 
242 	dma_attr = &sc->sge.rxb_params.dma_attr_rx;
243 	dma_attr->dma_attr_version = DMA_ATTR_V0;
244 	dma_attr->dma_attr_addr_lo = 0;
245 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
246 	dma_attr->dma_attr_count_max = UINT64_MAX;
247 	/*
248 	 * Low 4 bits of an rx buffer address have a special meaning to the SGE
249 	 * and an rx buf cannot have an address with any of these bits set.
250 	 * FL_ALIGN is >= 32 so we're sure things are ok.
251 	 */
252 	dma_attr->dma_attr_align = FL_ALIGN;
253 	dma_attr->dma_attr_burstsizes = 0xfff;
254 	dma_attr->dma_attr_minxfer = 1;
255 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
256 	dma_attr->dma_attr_seg = UINT64_MAX;
257 	dma_attr->dma_attr_sgllen = 1;
258 	dma_attr->dma_attr_granular = 1;
259 	dma_attr->dma_attr_flags = 0;
260 
261 	sc->sge.rxbuf_cache = rxbuf_cache_create(&sc->sge.rxb_params);
262 
263 	v = t4_read_reg(sc, A_SGE_CONM_CTRL);
264 	sc->sge.fl_starve_threshold = G_EGRTHRESHOLD(v) * 2 + 1;
265 
266 	t4_set_reg_field(sc, A_SGE_CONTROL, V_PKTSHIFT(M_PKTSHIFT) |
267 	    V_INGPADBOUNDARY(M_INGPADBOUNDARY) |
268 	    F_EGRSTATUSPAGESIZE, V_INGPADBOUNDARY(ilog2(FL_ALIGN) - 5) |
269 	    V_PKTSHIFT(FL_PKTSHIFT) |
270 	    F_RXPKTCPLMODE |
271 	    V_EGRSTATUSPAGESIZE(SPG_SIZE == 128));
272 
273 	t4_set_reg_field(sc, A_SGE_HOST_PAGE_SIZE,
274 	    V_HOSTPAGESIZEPF0(M_HOSTPAGESIZEPF0),
275 	    V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10));
276 
277 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0, rx_buf_size);
278 
279 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD,
280 	    V_THRESHOLD_0(p->counter_val[0]) |
281 	    V_THRESHOLD_1(p->counter_val[1]) |
282 	    V_THRESHOLD_2(p->counter_val[2]) |
283 	    V_THRESHOLD_3(p->counter_val[3]));
284 
285 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1,
286 	    V_TIMERVALUE0(us_to_core_ticks(sc, p->timer_val[0])) |
287 	    V_TIMERVALUE1(us_to_core_ticks(sc, p->timer_val[1])));
288 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3,
289 	    V_TIMERVALUE2(us_to_core_ticks(sc, p->timer_val[2])) |
290 	    V_TIMERVALUE3(us_to_core_ticks(sc, p->timer_val[3])));
291 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5,
292 	    V_TIMERVALUE4(us_to_core_ticks(sc, p->timer_val[4])) |
293 	    V_TIMERVALUE5(us_to_core_ticks(sc, p->timer_val[5])));
294 
295 	(void) t4_register_cpl_handler(sc, CPL_FW4_MSG, handle_fw_rpl);
296 	(void) t4_register_cpl_handler(sc, CPL_FW6_MSG, handle_fw_rpl);
297 	(void) t4_register_cpl_handler(sc, CPL_RX_PKT, t4_eth_rx);
298 }
299 
300 /*
301  * Allocate and initialize the firmware event queue and the forwarded interrupt
302  * queues, if any.  The adapter owns all these queues as they are not associated
303  * with any particular port.
304  *
305  * Returns errno on failure.  Resources allocated up to that point may still be
306  * allocated.  Caller is responsible for cleanup in case this function fails.
307  */
308 int
309 t4_setup_adapter_queues(struct adapter *sc)
310 {
311 	int rc;
312 
313 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
314 
315 	/*
316 	 * Firmware event queue
317 	 */
318 	rc = alloc_fwq(sc);
319 	if (rc != 0)
320 		return (rc);
321 
322 	/*
323 	 * Management queue.  This is just a control queue that uses the fwq as
324 	 * its associated iq.
325 	 */
326 	rc = alloc_mgmtq(sc);
327 
328 	return (rc);
329 }
330 
331 /*
332  * Idempotent
333  */
334 int
335 t4_teardown_adapter_queues(struct adapter *sc)
336 {
337 
338 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
339 
340 	(void) free_fwq(sc);
341 
342 	return (0);
343 }
344 
345 static inline int
346 first_vector(struct port_info *pi)
347 {
348 	struct adapter *sc = pi->adapter;
349 	int rc = T4_EXTRA_INTR, i;
350 
351 	if (sc->intr_count == 1)
352 		return (0);
353 
354 	for_each_port(sc, i) {
355 		struct port_info *p = sc->port[i];
356 
357 		if (i == pi->port_id)
358 			break;
359 
360 #ifndef TCP_OFFLOAD_DISABLE
361 		if (!(sc->flags & INTR_FWD))
362 			rc += p->nrxq + p->nofldrxq;
363 		else
364 			rc += max(p->nrxq, p->nofldrxq);
365 #else
366 		/*
367 		 * Not compiled with offload support and intr_count > 1.  Only
368 		 * NIC queues exist and they'd better be taking direct
369 		 * interrupts.
370 		 */
371 		KASSERT(!(sc->flags & INTR_FWD));
372 		rc += p->nrxq;
373 #endif
374 	}
375 	return (rc);
376 }
377 
378 /*
379  * Given an arbitrary "index," come up with an iq that can be used by other
380  * queues (of this port) for interrupt forwarding, SGE egress updates, etc.
381  * The iq returned is guaranteed to be something that takes direct interrupts.
382  */
383 static struct sge_iq *
384 port_intr_iq(struct port_info *pi, int idx)
385 {
386 	struct adapter *sc = pi->adapter;
387 	struct sge *s = &sc->sge;
388 	struct sge_iq *iq = NULL;
389 
390 	if (sc->intr_count == 1)
391 		return (&sc->sge.fwq);
392 
393 #ifndef TCP_OFFLOAD_DISABLE
394 	if (!(sc->flags & INTR_FWD)) {
395 		idx %= pi->nrxq + pi->nofldrxq;
396 
397 		if (idx >= pi->nrxq) {
398 			idx -= pi->nrxq;
399 			iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq;
400 		} else
401 			iq = &s->rxq[pi->first_rxq + idx].iq;
402 
403 	} else {
404 		idx %= max(pi->nrxq, pi->nofldrxq);
405 
406 		if (pi->nrxq >= pi->nofldrxq)
407 			iq = &s->rxq[pi->first_rxq + idx].iq;
408 		else
409 			iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq;
410 	}
411 #else
412 	/*
413 	 * Not compiled with offload support and intr_count > 1.  Only NIC
414 	 * queues exist and they'd better be taking direct interrupts.
415 	 */
416 	ASSERT(!(sc->flags & INTR_FWD));
417 
418 	idx %= pi->nrxq;
419 	iq = &s->rxq[pi->first_rxq + idx].iq;
420 #endif
421 
422 	return (iq);
423 }
424 
425 int
426 t4_setup_port_queues(struct port_info *pi)
427 {
428 	int rc = 0, i, intr_idx, j, iqid;
429 	struct sge_rxq *rxq;
430 	struct sge_txq *txq;
431 #ifndef TCP_OFFLOAD_DISABLE
432 	struct sge_wrq *ctrlq;
433 	struct sge_ofld_rxq *ofld_rxq;
434 	struct sge_wrq *ofld_txq;
435 #endif
436 	struct adapter *sc = pi->adapter;
437 	struct driver_properties *p = &sc->props;
438 
439 	pi->ksp_config = setup_port_config_kstats(pi);
440 	pi->ksp_info   = setup_port_info_kstats(pi);
441 
442 	/* Interrupt vector to start from (when using multiple vectors) */
443 	intr_idx = first_vector(pi);
444 
445 	/*
446 	 * First pass over all rx queues (NIC and TOE):
447 	 * a) initialize iq and fl
448 	 * b) allocate queue iff it will take direct interrupts.
449 	 */
450 
451 	for_each_rxq(pi, i, rxq) {
452 
453 		init_iq(&rxq->iq, sc, pi->tmr_idx, pi->pktc_idx, p->qsize_rxq,
454 		    RX_IQ_ESIZE);
455 
456 		init_fl(&rxq->fl, p->qsize_rxq / 8); /* 8 bufs in each entry */
457 
458 #ifndef TCP_OFFLOAD_DISABLE
459 		if ((!(sc->flags & INTR_FWD)) ||
460 		    (sc->intr_count > 1 && pi->nrxq >= pi->nofldrxq))
461 #else
462 		if (!(sc->flags & INTR_FWD))
463 #endif
464 		{
465 			rxq->iq.flags |= IQ_INTR;
466 			rc = alloc_rxq(pi, rxq, intr_idx, i);
467 			if (rc != 0)
468 				goto done;
469 			intr_idx++;
470 		}
471 
472 	}
473 
474 #ifndef TCP_OFFLOAD_DISABLE
475 	for_each_ofld_rxq(pi, i, ofld_rxq) {
476 
477 		init_iq(&ofld_rxq->iq, sc, pi->tmr_idx, pi->pktc_idx,
478 		    p->qsize_rxq, RX_IQ_ESIZE);
479 
480 		init_fl(&ofld_rxq->fl, p->qsize_rxq / 8);
481 
482 		if (!(sc->flags & INTR_FWD) ||
483 		    (sc->intr_count > 1 && pi->nofldrxq > pi->nrxq)) {
484 			ofld_rxq->iq.flags = IQ_INTR;
485 			rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx);
486 			if (rc != 0)
487 				goto done;
488 
489 			intr_idx++;
490 		}
491 	}
492 #endif
493 
494 	/*
495 	 * Second pass over all rx queues (NIC and TOE).  The queues forwarding
496 	 * their interrupts are allocated now.
497 	 */
498 	j = 0;
499 	for_each_rxq(pi, i, rxq) {
500 		if (rxq->iq.flags & IQ_INTR)
501 			continue;
502 
503 		intr_idx = port_intr_iq(pi, j)->abs_id;
504 
505 		rc = alloc_rxq(pi, rxq, intr_idx, i);
506 		if (rc != 0)
507 			goto done;
508 		j++;
509 	}
510 
511 #ifndef TCP_OFFLOAD_DISABLE
512 	for_each_ofld_rxq(pi, i, ofld_rxq) {
513 		if (ofld_rxq->iq.flags & IQ_INTR)
514 			continue;
515 
516 		intr_idx = port_intr_iq(pi, j)->abs_id;
517 		rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx);
518 		if (rc != 0)
519 			goto done;
520 		j++;
521 	}
522 #endif
523 	/*
524 	 * Now the tx queues.  Only one pass needed.
525 	 */
526 	j = 0;
527 	for_each_txq(pi, i, txq) {
528 		uint16_t iqid;
529 
530 		iqid = port_intr_iq(pi, j)->cntxt_id;
531 		init_eq(&txq->eq, EQ_ETH, p->qsize_txq, pi->tx_chan, iqid);
532 		rc = alloc_txq(pi, txq, i);
533 		if (rc != 0)
534 			goto done;
535 	}
536 
537 #ifndef TCP_OFFLOAD_DISABLE
538 	for_each_ofld_txq(pi, i, ofld_txq) {
539 		uint16_t iqid;
540 
541 		iqid = port_intr_iq(pi, j)->cntxt_id;
542 		init_eq(&ofld_txq->eq, EQ_OFLD, p->qsize_txq, pi->tx_chan,
543 		    iqid);
544 		rc = alloc_wrq(sc, pi, ofld_txq, i);
545 		if (rc != 0)
546 			goto done;
547 	}
548 #endif
549 
550 	/*
551 	 * Finally, the control queue.
552 	 */
553 	ctrlq = &sc->sge.ctrlq[pi->port_id];
554 	iqid = port_intr_iq(pi, 0)->cntxt_id;
555 	init_eq(&ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid);
556 	rc = alloc_wrq(sc, pi, ctrlq, 0);
557 
558 done:
559 	if (rc != 0)
560 		(void) t4_teardown_port_queues(pi);
561 
562 	return (rc);
563 }
564 
565 /*
566  * Idempotent
567  */
568 int
569 t4_teardown_port_queues(struct port_info *pi)
570 {
571 	int i;
572 	struct sge_rxq *rxq;
573 	struct sge_txq *txq;
574 	struct adapter *sc = pi->adapter;
575 #ifndef TCP_OFFLOAD_DISABLE
576 	struct sge_ofld_rxq *ofld_rxq;
577 	struct sge_wrq *ofld_txq;
578 #endif
579 
580 	if (pi->ksp_config != NULL) {
581 		kstat_delete(pi->ksp_config);
582 		pi->ksp_config = NULL;
583 	}
584 	if (pi->ksp_info != NULL) {
585 		kstat_delete(pi->ksp_info);
586 		pi->ksp_info = NULL;
587 	}
588 
589 	(void) free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
590 
591 	for_each_txq(pi, i, txq) {
592 		(void) free_txq(pi, txq);
593 	}
594 
595 #ifndef TCP_OFFLOAD_DISABLE
596 	for_each_ofld_txq(pi, i, ofld_txq) {
597 		(void) free_wrq(sc, ofld_txq);
598 	}
599 
600 	for_each_ofld_rxq(pi, i, ofld_rxq) {
601 		if ((ofld_rxq->iq.flags & IQ_INTR) == 0)
602 			(void) free_ofld_rxq(pi, ofld_rxq);
603 	}
604 #endif
605 
606 	for_each_rxq(pi, i, rxq) {
607 		if ((rxq->iq.flags & IQ_INTR) == 0)
608 			(void) free_rxq(pi, rxq);
609 	}
610 
611 	/*
612 	 * Then take down the rx queues that take direct interrupts.
613 	 */
614 
615 	for_each_rxq(pi, i, rxq) {
616 		if (rxq->iq.flags & IQ_INTR)
617 			(void) free_rxq(pi, rxq);
618 	}
619 
620 #ifndef TCP_OFFLOAD_DISABLE
621 	for_each_ofld_rxq(pi, i, ofld_rxq) {
622 		if (ofld_rxq->iq.flags & IQ_INTR)
623 			(void) free_ofld_rxq(pi, ofld_rxq);
624 	}
625 #endif
626 
627 	return (0);
628 }
629 
630 /* Deals with errors and forwarded interrupts */
631 uint_t
632 t4_intr_all(caddr_t arg1, caddr_t arg2)
633 {
634 
635 	(void) t4_intr_err(arg1, arg2);
636 	(void) t4_intr(arg1, arg2);
637 
638 	return (DDI_INTR_CLAIMED);
639 }
640 
641 /* Deals with interrupts on the given ingress queue */
642 /* ARGSUSED */
643 uint_t
644 t4_intr(caddr_t arg1, caddr_t arg2)
645 {
646 	/* LINTED: E_BAD_PTR_CAST_ALIGN */
647 	struct sge_iq *iq = (struct sge_iq *)arg2;
648 
649 	if (atomic_cas_uint(&iq->state, IQS_IDLE, IQS_BUSY) == IQS_IDLE) {
650 		(void) service_iq(iq, 0);
651 		(void) atomic_cas_uint(&iq->state, IQS_BUSY, IQS_IDLE);
652 	}
653 	return (DDI_INTR_CLAIMED);
654 }
655 
656 /* Deals with error interrupts */
657 /* ARGSUSED */
658 uint_t
659 t4_intr_err(caddr_t arg1, caddr_t arg2)
660 {
661 	/* LINTED: E_BAD_PTR_CAST_ALIGN */
662 	struct adapter *sc = (struct adapter *)arg1;
663 
664 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
665 	(void) t4_slow_intr_handler(sc);
666 
667 	return (DDI_INTR_CLAIMED);
668 }
669 
670 /*
671  * Deals with anything and everything on the given ingress queue.
672  */
673 static int
674 service_iq(struct sge_iq *iq, int budget)
675 {
676 	struct sge_iq *q;
677 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
678 	struct sge_fl *fl = &rxq->fl;		/* Use iff IQ_HAS_FL */
679 	struct adapter *sc = iq->adapter;
680 	struct rsp_ctrl *ctrl;
681 	const struct rss_header *rss;
682 	int ndescs = 0, limit, fl_bufs_used = 0;
683 	int rsp_type;
684 	uint32_t lq;
685 	mblk_t *m;
686 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
687 
688 	limit = budget ? budget : iq->qsize / 8;
689 
690 	/*
691 	 * We always come back and check the descriptor ring for new indirect
692 	 * interrupts and other responses after running a single handler.
693 	 */
694 	for (;;) {
695 		while (is_new_response(iq, &ctrl)) {
696 
697 			membar_consumer();
698 
699 			m = NULL;
700 			rsp_type = G_RSPD_TYPE(ctrl->u.type_gen);
701 			lq = be32_to_cpu(ctrl->pldbuflen_qid);
702 			rss = (const void *)iq->cdesc;
703 
704 			switch (rsp_type) {
705 			case X_RSPD_TYPE_FLBUF:
706 
707 				ASSERT(iq->flags & IQ_HAS_FL);
708 
709 				m = get_fl_payload(fl, lq, &fl_bufs_used);
710 				if (m == NULL) {
711 					panic("%s: line %d.", __func__,
712 					    __LINE__);
713 				}
714 
715 			/* FALLTHRU */
716 			case X_RSPD_TYPE_CPL:
717 
718 				ASSERT(rss->opcode < NUM_CPL_CMDS);
719 				sc->cpl_handler[rss->opcode](iq, rss, m);
720 				break;
721 
722 			case X_RSPD_TYPE_INTR:
723 
724 				/*
725 				 * Interrupts should be forwarded only to queues
726 				 * that are not forwarding their interrupts.
727 				 * This means service_iq can recurse but only 1
728 				 * level deep.
729 				 */
730 				ASSERT(budget == 0);
731 
732 				q = sc->sge.iqmap[lq - sc->sge.iq_start];
733 				if (atomic_cas_uint(&q->state, IQS_IDLE,
734 				    IQS_BUSY) == IQS_IDLE) {
735 					if (service_iq(q, q->qsize / 8) == 0) {
736 						(void) atomic_cas_uint(
737 						    &q->state, IQS_BUSY,
738 						    IQS_IDLE);
739 					} else {
740 						STAILQ_INSERT_TAIL(&iql, q,
741 						    link);
742 					}
743 				}
744 				break;
745 
746 			default:
747 				break;
748 			}
749 
750 			iq_next(iq);
751 			if (++ndescs == limit) {
752 				t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
753 				    V_CIDXINC(ndescs) |
754 				    V_INGRESSQID(iq->cntxt_id) |
755 				    V_SEINTARM(V_QINTR_TIMER_IDX(
756 				    X_TIMERREG_UPDATE_CIDX)));
757 				ndescs = 0;
758 
759 				if (fl_bufs_used > 0) {
760 					ASSERT(iq->flags & IQ_HAS_FL);
761 					FL_LOCK(fl);
762 					fl->needed += fl_bufs_used;
763 					(void) refill_fl(sc, fl, fl->cap / 8);
764 					FL_UNLOCK(fl);
765 					fl_bufs_used = 0;
766 				}
767 
768 				if (budget != 0)
769 					return (EINPROGRESS);
770 			}
771 		}
772 
773 		if (STAILQ_EMPTY(&iql) != 0)
774 			break;
775 
776 		/*
777 		 * Process the head only, and send it to the back of the list if
778 		 * it's still not done.
779 		 */
780 		q = STAILQ_FIRST(&iql);
781 		STAILQ_REMOVE_HEAD(&iql, link);
782 		if (service_iq(q, q->qsize / 8) == 0)
783 			(void) atomic_cas_uint(&q->state, IQS_BUSY, IQS_IDLE);
784 		else
785 			STAILQ_INSERT_TAIL(&iql, q, link);
786 	}
787 
788 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
789 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_next));
790 
791 	if (iq->flags & IQ_HAS_FL) {
792 		int starved;
793 
794 		FL_LOCK(fl);
795 		fl->needed += fl_bufs_used;
796 		starved = refill_fl(sc, fl, fl->cap / 4);
797 		FL_UNLOCK(fl);
798 		if (starved != 0)
799 			add_fl_to_sfl(sc, fl);
800 	}
801 
802 	return (0);
803 }
804 
805 int
806 t4_mgmt_tx(struct adapter *sc, mblk_t *m)
807 {
808 	return (t4_wrq_tx(sc, &sc->sge.mgmtq, m));
809 }
810 
811 /*
812  * Doesn't fail.  Holds on to work requests it can't send right away.
813  */
814 int
815 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, mblk_t *m0)
816 {
817 	struct sge_eq *eq = &wrq->eq;
818 	struct mblk_pair *wr_list = &wrq->wr_list;
819 	int can_reclaim;
820 	caddr_t dst;
821 	mblk_t *wr, *next;
822 
823 	TXQ_LOCK_ASSERT_OWNED(wrq);
824 #ifndef TCP_OFFLOAD_DISABLE
825 	ASSERT((eq->flags & EQ_TYPEMASK) == EQ_OFLD ||
826 	    (eq->flags & EQ_TYPEMASK) == EQ_CTRL);
827 #else
828 	ASSERT((eq->flags & EQ_TYPEMASK) == EQ_CTRL);
829 #endif
830 
831 	if (m0 != NULL) {
832 		if (wr_list->head != NULL)
833 			wr_list->tail->b_next = m0;
834 		else
835 			wr_list->head = m0;
836 		while (m0->b_next)
837 			m0 = m0->b_next;
838 		wr_list->tail = m0;
839 	}
840 
841 	can_reclaim = reclaimable(eq);
842 	eq->cidx += can_reclaim;
843 	eq->avail += can_reclaim;
844 	if (eq->cidx >= eq->cap)
845 		eq->cidx -= eq->cap;
846 
847 	for (wr = wr_list->head; wr; wr = next) {
848 		int ndesc, len = 0;
849 		mblk_t *m;
850 
851 		next = wr->b_next;
852 		wr->b_next = NULL;
853 
854 		for (m = wr; m; m = m->b_cont)
855 			len += MBLKL(m);
856 
857 		ASSERT(len > 0 && (len & 0x7) == 0);
858 		ASSERT(len <= SGE_MAX_WR_LEN);
859 
860 		ndesc = howmany(len, EQ_ESIZE);
861 		if (eq->avail < ndesc) {
862 			wr->b_next = next;
863 			wrq->no_desc++;
864 			break;
865 		}
866 
867 		dst = (void *)&eq->desc[eq->pidx];
868 		for (m = wr; m; m = m->b_cont)
869 			copy_to_txd(eq, (void *)m->b_rptr, &dst, MBLKL(m));
870 
871 		eq->pidx += ndesc;
872 		eq->avail -= ndesc;
873 		if (eq->pidx >= eq->cap)
874 			eq->pidx -= eq->cap;
875 
876 		eq->pending += ndesc;
877 		if (eq->pending > 16)
878 			ring_tx_db(sc, eq);
879 
880 		wrq->tx_wrs++;
881 		freemsg(wr);
882 
883 		if (eq->avail < 8) {
884 			can_reclaim = reclaimable(eq);
885 			eq->cidx += can_reclaim;
886 			eq->avail += can_reclaim;
887 			if (eq->cidx >= eq->cap)
888 				eq->cidx -= eq->cap;
889 		}
890 	}
891 
892 	if (eq->pending != 0)
893 		ring_tx_db(sc, eq);
894 
895 	if (wr == NULL)
896 		wr_list->head = wr_list->tail = NULL;
897 	else {
898 		wr_list->head = wr;
899 
900 		ASSERT(wr_list->tail->b_next == NULL);
901 	}
902 
903 	return (0);
904 }
905 
906 /* Per-packet header in a coalesced tx WR, before the SGL starts (in flits) */
907 #define	TXPKTS_PKT_HDR ((\
908 	sizeof (struct ulp_txpkt) + \
909 	sizeof (struct ulptx_idata) + \
910 	sizeof (struct cpl_tx_pkt_core)) / 8)
911 
912 /* Header of a coalesced tx WR, before SGL of first packet (in flits) */
913 #define	TXPKTS_WR_HDR (\
914 	sizeof (struct fw_eth_tx_pkts_wr) / 8 + \
915 	TXPKTS_PKT_HDR)
916 
917 /* Header of a tx WR, before SGL of first packet (in flits) */
918 #define	TXPKT_WR_HDR ((\
919 	sizeof (struct fw_eth_tx_pkt_wr) + \
920 	sizeof (struct cpl_tx_pkt_core)) / 8)
921 
922 /* Header of a tx LSO WR, before SGL of first packet (in flits) */
923 #define	TXPKT_LSO_WR_HDR ((\
924 	sizeof (struct fw_eth_tx_pkt_wr) + \
925 	sizeof (struct cpl_tx_pkt_lso) + \
926 	sizeof (struct cpl_tx_pkt_core)) / 8)
927 
928 mblk_t *
929 t4_eth_tx(struct port_info *pi, struct sge_txq *txq, mblk_t *frame)
930 {
931 	struct adapter *sc = pi->adapter;
932 	struct sge_eq *eq = &txq->eq;
933 	mblk_t *next_frame;
934 	int rc, coalescing;
935 	struct txpkts txpkts;
936 	struct txinfo txinfo;
937 
938 	txpkts.npkt = 0; /* indicates there's nothing in txpkts */
939 	coalescing = 0;
940 
941 	TXQ_LOCK(txq);
942 	if (eq->avail < 8)
943 		(void) reclaim_tx_descs(txq, 8);
944 	for (; frame; frame = next_frame) {
945 
946 		if (eq->avail < 8)
947 			break;
948 
949 		next_frame = frame->b_next;
950 		frame->b_next = NULL;
951 
952 		if (next_frame != NULL)
953 			coalescing = 1;
954 
955 		rc = get_frame_txinfo(txq, &frame, &txinfo, coalescing);
956 		if (rc != 0) {
957 			if (rc == ENOMEM) {
958 
959 				/* Short of resources, suspend tx */
960 
961 				frame->b_next = next_frame;
962 				break;
963 			}
964 
965 			/*
966 			 * Unrecoverable error for this frame, throw it
967 			 * away and move on to the next.
968 			 */
969 
970 			freemsg(frame);
971 			continue;
972 		}
973 
974 		if (coalescing != 0 &&
975 		    add_to_txpkts(txq, &txpkts, frame, &txinfo) == 0) {
976 
977 			/* Successfully absorbed into txpkts */
978 
979 			write_ulp_cpl_sgl(pi, txq, &txpkts, &txinfo);
980 			goto doorbell;
981 		}
982 
983 		/*
984 		 * We weren't coalescing to begin with, or current frame could
985 		 * not be coalesced (add_to_txpkts flushes txpkts if a frame
986 		 * given to it can't be coalesced).  Either way there should be
987 		 * nothing in txpkts.
988 		 */
989 		ASSERT(txpkts.npkt == 0);
990 
991 		/* We're sending out individual frames now */
992 		coalescing = 0;
993 
994 		if (eq->avail < 8)
995 			(void) reclaim_tx_descs(txq, 8);
996 		rc = write_txpkt_wr(pi, txq, frame, &txinfo);
997 		if (rc != 0) {
998 
999 			/* Short of hardware descriptors, suspend tx */
1000 
1001 			/*
1002 			 * This is an unlikely but expensive failure.  We've
1003 			 * done all the hard work (DMA bindings etc.) and now we
1004 			 * can't send out the frame.  What's worse, we have to
1005 			 * spend even more time freeing up everything in txinfo.
1006 			 */
1007 			txq->qfull++;
1008 			free_txinfo_resources(txq, &txinfo);
1009 
1010 			frame->b_next = next_frame;
1011 			break;
1012 		}
1013 
1014 doorbell:
1015 		/* Fewer and fewer doorbells as the queue fills up */
1016 		if (eq->pending >= (1 << (fls(eq->qsize - eq->avail) / 2)))
1017 			ring_tx_db(sc, eq);
1018 		(void) reclaim_tx_descs(txq, 32);
1019 	}
1020 
1021 	if (txpkts.npkt > 0)
1022 		write_txpkts_wr(txq, &txpkts);
1023 
1024 	/*
1025 	 * frame not NULL means there was an error but we haven't thrown it
1026 	 * away.  This can happen when we're short of tx descriptors (qfull) or
1027 	 * maybe even DMA handles (dma_hdl_failed).  Either way, a credit flush
1028 	 * and reclaim will get things going again.
1029 	 *
1030 	 * If eq->avail is already 0 we know a credit flush was requested in the
1031 	 * WR that reduced it to 0 so we don't need another flush (we don't have
1032 	 * any descriptor for a flush WR anyway, duh).
1033 	 */
1034 	if (frame && eq->avail > 0)
1035 		write_txqflush_wr(txq);
1036 
1037 	if (eq->pending != 0)
1038 		ring_tx_db(sc, eq);
1039 
1040 	(void) reclaim_tx_descs(txq, eq->qsize);
1041 	TXQ_UNLOCK(txq);
1042 
1043 	return (frame);
1044 }
1045 
1046 static inline void
1047 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int8_t pktc_idx, int
1048     qsize, uint8_t esize)
1049 {
1050 	ASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS);
1051 	ASSERT(pktc_idx < SGE_NCOUNTERS);	/* -ve is ok, means don't use */
1052 
1053 	iq->flags = 0;
1054 	iq->adapter = sc;
1055 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
1056 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
1057 	if (pktc_idx >= 0) {
1058 		iq->intr_params |= F_QINTR_CNT_EN;
1059 		iq->intr_pktc_idx = pktc_idx;
1060 	}
1061 	iq->qsize = roundup(qsize, 16);		/* See FW_IQ_CMD/iqsize */
1062 	iq->esize = max(esize, 16);		/* See FW_IQ_CMD/iqesize */
1063 }
1064 
1065 static inline void
1066 init_fl(struct sge_fl *fl, uint16_t qsize)
1067 {
1068 
1069 	fl->qsize = qsize;
1070 }
1071 
1072 static inline void
1073 init_eq(struct sge_eq *eq, uint16_t eqtype, uint16_t qsize, uint8_t tx_chan,
1074     uint16_t iqid)
1075 {
1076 	ASSERT(tx_chan < NCHAN);
1077 	ASSERT(eqtype <= EQ_TYPEMASK);
1078 
1079 	eq->flags = eqtype & EQ_TYPEMASK;
1080 	eq->tx_chan = tx_chan;
1081 	eq->iqid = iqid;
1082 	eq->qsize = qsize;
1083 }
1084 
1085 /*
1086  * Allocates the ring for an ingress queue and an optional freelist.  If the
1087  * freelist is specified it will be allocated and then associated with the
1088  * ingress queue.
1089  *
1090  * Returns errno on failure.  Resources allocated up to that point may still be
1091  * allocated.  Caller is responsible for cleanup in case this function fails.
1092  *
1093  * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
1094  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
1095  * the index of the queue to which its interrupts will be forwarded.
1096  */
1097 static int
1098 alloc_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl,
1099     int intr_idx, int cong)
1100 {
1101 	int rc, cntxt_id;
1102 	size_t len;
1103 	struct fw_iq_cmd c;
1104 	struct adapter *sc = iq->adapter;
1105 	uint32_t v = 0;
1106 
1107 	len = iq->qsize * iq->esize;
1108 	rc = alloc_desc_ring(sc, len, DDI_DMA_READ, &iq->dhdl, &iq->ahdl,
1109 	    &iq->ba, (caddr_t *)&iq->desc);
1110 	if (rc != 0)
1111 		return (rc);
1112 
1113 	bzero(&c, sizeof (c));
1114 	c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
1115 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
1116 	    V_FW_IQ_CMD_VFN(0));
1117 
1118 	c.alloc_to_len16 = cpu_to_be32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
1119 	    FW_LEN16(c));
1120 
1121 	/* Special handling for firmware event queue */
1122 	if (iq == &sc->sge.fwq)
1123 		v |= F_FW_IQ_CMD_IQASYNCH;
1124 
1125 	if (iq->flags & IQ_INTR)
1126 		ASSERT(intr_idx < sc->intr_count);
1127 	else
1128 		v |= F_FW_IQ_CMD_IQANDST;
1129 	v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
1130 
1131 	c.type_to_iqandstindex = cpu_to_be32(v |
1132 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
1133 	    V_FW_IQ_CMD_VIID(pi->viid) |
1134 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
1135 	c.iqdroprss_to_iqesize = cpu_to_be16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
1136 	    F_FW_IQ_CMD_IQGTSMODE |
1137 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
1138 	    V_FW_IQ_CMD_IQESIZE(ilog2(iq->esize) - 4));
1139 	c.iqsize = cpu_to_be16(iq->qsize);
1140 	c.iqaddr = cpu_to_be64(iq->ba);
1141 	if (cong >= 0)
1142 		c.iqns_to_fl0congen = BE_32(F_FW_IQ_CMD_IQFLINTCONGEN);
1143 
1144 	if (fl != NULL) {
1145 		mutex_init(&fl->lock, NULL, MUTEX_DRIVER,
1146 		    DDI_INTR_PRI(sc->intr_pri));
1147 		fl->flags |= FL_MTX;
1148 
1149 		len = fl->qsize * RX_FL_ESIZE;
1150 		rc = alloc_desc_ring(sc, len, DDI_DMA_WRITE, &fl->dhdl,
1151 		    &fl->ahdl, &fl->ba, (caddr_t *)&fl->desc);
1152 		if (rc != 0)
1153 			return (rc);
1154 
1155 		/* Allocate space for one software descriptor per buffer. */
1156 		fl->cap = (fl->qsize - SPG_SIZE / RX_FL_ESIZE) * 8;
1157 		fl->sdesc = kmem_zalloc(sizeof (struct fl_sdesc) * fl->cap,
1158 		    KM_SLEEP);
1159 		fl->needed = fl->cap;
1160 		fl->lowat = roundup(sc->sge.fl_starve_threshold, 8);
1161 
1162 		c.iqns_to_fl0congen |=
1163 		    cpu_to_be32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
1164 		    F_FW_IQ_CMD_FL0PACKEN | F_FW_IQ_CMD_FL0PADEN);
1165 		if (cong >= 0) {
1166 			c.iqns_to_fl0congen |=
1167 			    BE_32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
1168 			    F_FW_IQ_CMD_FL0CONGCIF |
1169 			    F_FW_IQ_CMD_FL0CONGEN);
1170 		}
1171 		c.fl0dcaen_to_fl0cidxfthresh =
1172 		    cpu_to_be16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_64B) |
1173 		    V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B));
1174 		c.fl0size = cpu_to_be16(fl->qsize);
1175 		c.fl0addr = cpu_to_be64(fl->ba);
1176 	}
1177 
1178 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1179 	if (rc != 0) {
1180 		cxgb_printf(sc->dip, CE_WARN,
1181 		    "failed to create ingress queue: %d", rc);
1182 		return (rc);
1183 	}
1184 
1185 	iq->cdesc = iq->desc;
1186 	iq->cidx = 0;
1187 	iq->gen = 1;
1188 	iq->intr_next = iq->intr_params;
1189 	iq->adapter = sc;
1190 	iq->cntxt_id = be16_to_cpu(c.iqid);
1191 	iq->abs_id = be16_to_cpu(c.physiqid);
1192 	iq->flags |= IQ_ALLOCATED;
1193 
1194 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
1195 	if (cntxt_id >= sc->sge.niq) {
1196 		panic("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
1197 		    cntxt_id, sc->sge.niq - 1);
1198 	}
1199 	sc->sge.iqmap[cntxt_id] = iq;
1200 
1201 	if (fl != NULL) {
1202 		fl->cntxt_id = be16_to_cpu(c.fl0id);
1203 		fl->pidx = fl->cidx = 0;
1204 		fl->copy_threshold = rx_copy_threshold;
1205 
1206 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
1207 		if (cntxt_id >= sc->sge.neq) {
1208 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
1209 			    __func__, cntxt_id, sc->sge.neq - 1);
1210 		}
1211 		sc->sge.eqmap[cntxt_id] = (void *)fl;
1212 
1213 		FL_LOCK(fl);
1214 		(void) refill_fl(sc, fl, fl->lowat);
1215 		FL_UNLOCK(fl);
1216 
1217 		iq->flags |= IQ_HAS_FL;
1218 	}
1219 
1220 	/* Enable IQ interrupts */
1221 	iq->state = IQS_IDLE;
1222 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_SEINTARM(iq->intr_params) |
1223 	    V_INGRESSQID(iq->cntxt_id));
1224 
1225 	return (0);
1226 }
1227 
1228 static int
1229 free_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl)
1230 {
1231 	int rc;
1232 	struct adapter *sc = iq->adapter;
1233 	dev_info_t *dip;
1234 
1235 	dip = pi ? pi->dip : sc->dip;
1236 
1237 	if (iq != NULL) {
1238 		if (iq->flags & IQ_ALLOCATED) {
1239 			rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
1240 			    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
1241 			    fl ? fl->cntxt_id : 0xffff, 0xffff);
1242 			if (rc != 0) {
1243 				cxgb_printf(dip, CE_WARN,
1244 				    "failed to free queue %p: %d", iq, rc);
1245 				return (rc);
1246 			}
1247 			iq->flags &= ~IQ_ALLOCATED;
1248 		}
1249 
1250 		if (iq->desc != NULL) {
1251 			(void) free_desc_ring(&iq->dhdl, &iq->ahdl);
1252 			iq->desc = NULL;
1253 		}
1254 
1255 		bzero(iq, sizeof (*iq));
1256 	}
1257 
1258 	if (fl != NULL) {
1259 		if (fl->sdesc != NULL) {
1260 			FL_LOCK(fl);
1261 			free_fl_bufs(fl);
1262 			FL_UNLOCK(fl);
1263 
1264 			kmem_free(fl->sdesc, sizeof (struct fl_sdesc) *
1265 			    fl->cap);
1266 			fl->sdesc = NULL;
1267 		}
1268 
1269 		if (fl->desc != NULL) {
1270 			(void) free_desc_ring(&fl->dhdl, &fl->ahdl);
1271 			fl->desc = NULL;
1272 		}
1273 
1274 		if (fl->flags & FL_MTX) {
1275 			mutex_destroy(&fl->lock);
1276 			fl->flags &= ~FL_MTX;
1277 		}
1278 
1279 		bzero(fl, sizeof (struct sge_fl));
1280 	}
1281 
1282 	return (0);
1283 }
1284 
1285 static int
1286 alloc_fwq(struct adapter *sc)
1287 {
1288 	int rc, intr_idx;
1289 	struct sge_iq *fwq = &sc->sge.fwq;
1290 
1291 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE, FW_IQ_ESIZE);
1292 	fwq->flags |= IQ_INTR;	/* always */
1293 	intr_idx = sc->intr_count > 1 ? 1 : 0;
1294 	rc = alloc_iq_fl(sc->port[0], fwq, NULL, intr_idx, -1);
1295 	if (rc != 0) {
1296 		cxgb_printf(sc->dip, CE_WARN,
1297 		    "failed to create firmware event queue: %d.", rc);
1298 		return (rc);
1299 	}
1300 
1301 	return (0);
1302 }
1303 
1304 static int
1305 free_fwq(struct adapter *sc)
1306 {
1307 
1308 	return (free_iq_fl(NULL, &sc->sge.fwq, NULL));
1309 }
1310 
1311 static int
1312 alloc_mgmtq(struct adapter *sc)
1313 {
1314 	int rc;
1315 	struct sge_wrq *mgmtq = &sc->sge.mgmtq;
1316 
1317 	init_eq(&mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
1318 	    sc->sge.fwq.cntxt_id);
1319 	rc = alloc_wrq(sc, NULL, mgmtq, 0);
1320 	if (rc != 0) {
1321 		cxgb_printf(sc->dip, CE_WARN,
1322 		    "failed to create management queue: %d\n", rc);
1323 		return (rc);
1324 	}
1325 
1326 	return (0);
1327 }
1328 
1329 static int
1330 alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx, int i)
1331 {
1332 	int rc;
1333 
1334 	rxq->port = pi;
1335 	rc = alloc_iq_fl(pi, &rxq->iq, &rxq->fl, intr_idx, 1 << pi->tx_chan);
1336 	if (rc != 0)
1337 		return (rc);
1338 
1339 	rxq->ksp = setup_rxq_kstats(pi, rxq, i);
1340 
1341 	return (rc);
1342 }
1343 
1344 static int
1345 free_rxq(struct port_info *pi, struct sge_rxq *rxq)
1346 {
1347 	int rc;
1348 
1349 	if (rxq->ksp != NULL) {
1350 		kstat_delete(rxq->ksp);
1351 		rxq->ksp = NULL;
1352 	}
1353 
1354 	rc = free_iq_fl(pi, &rxq->iq, &rxq->fl);
1355 	if (rc == 0)
1356 		bzero(&rxq->fl, sizeof (*rxq) - offsetof(struct sge_rxq, fl));
1357 
1358 	return (rc);
1359 }
1360 
1361 #ifndef TCP_OFFLOAD_DISABLE
1362 static int
1363 alloc_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq,
1364 	int intr_idx)
1365 {
1366 	int rc;
1367 
1368 	rc = alloc_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx,
1369 	    1 << pi->tx_chan);
1370 	if (rc != 0)
1371 		return (rc);
1372 
1373 	return (rc);
1374 }
1375 
1376 static int
1377 free_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq)
1378 {
1379 	int rc;
1380 
1381 	rc = free_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl);
1382 	if (rc == 0)
1383 		bzero(&ofld_rxq->fl, sizeof (*ofld_rxq) -
1384 		    offsetof(struct sge_ofld_rxq, fl));
1385 
1386 	return (rc);
1387 }
1388 #endif
1389 
1390 static int
1391 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
1392 {
1393 	int rc, cntxt_id;
1394 	struct fw_eq_ctrl_cmd c;
1395 
1396 	bzero(&c, sizeof (c));
1397 
1398 	c.op_to_vfn = BE_32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
1399 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
1400 	    V_FW_EQ_CTRL_CMD_VFN(0));
1401 	c.alloc_to_len16 = BE_32(F_FW_EQ_CTRL_CMD_ALLOC |
1402 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
1403 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid)); /* TODO */
1404 	c.physeqid_pkd = BE_32(0);
1405 	c.fetchszm_to_iqid =
1406 	    BE_32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1407 	    V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
1408 	    F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
1409 	c.dcaen_to_eqsize =
1410 	    BE_32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1411 	    V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1412 	    V_FW_EQ_CTRL_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1413 	    V_FW_EQ_CTRL_CMD_EQSIZE(eq->qsize));
1414 	c.eqaddr = BE_64(eq->ba);
1415 
1416 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1417 	if (rc != 0) {
1418 		cxgb_printf(sc->dip, CE_WARN,
1419 		    "failed to create control queue %d: %d", eq->tx_chan, rc);
1420 		return (rc);
1421 	}
1422 	eq->flags |= EQ_ALLOCATED;
1423 
1424 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(BE_32(c.cmpliqid_eqid));
1425 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1426 	if (cntxt_id >= sc->sge.neq)
1427 		panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1428 		    cntxt_id, sc->sge.neq - 1);
1429 	sc->sge.eqmap[cntxt_id] = eq;
1430 
1431 	return (rc);
1432 }
1433 
1434 static int
1435 eth_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1436 {
1437 	int rc, cntxt_id;
1438 	struct fw_eq_eth_cmd c;
1439 
1440 	bzero(&c, sizeof (c));
1441 
1442 	c.op_to_vfn = BE_32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
1443 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
1444 	    V_FW_EQ_ETH_CMD_VFN(0));
1445 	c.alloc_to_len16 = BE_32(F_FW_EQ_ETH_CMD_ALLOC |
1446 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
1447 	c.viid_pkd = BE_32(V_FW_EQ_ETH_CMD_VIID(pi->viid));
1448 	c.fetchszm_to_iqid =
1449 	    BE_32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1450 	    V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
1451 	    V_FW_EQ_ETH_CMD_IQID(eq->iqid));
1452 	c.dcaen_to_eqsize = BE_32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1453 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1454 	    V_FW_EQ_ETH_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1455 	    V_FW_EQ_ETH_CMD_EQSIZE(eq->qsize));
1456 	c.eqaddr = BE_64(eq->ba);
1457 
1458 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1459 	if (rc != 0) {
1460 		cxgb_printf(pi->dip, CE_WARN,
1461 		    "failed to create Ethernet egress queue: %d", rc);
1462 		return (rc);
1463 	}
1464 	eq->flags |= EQ_ALLOCATED;
1465 
1466 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(BE_32(c.eqid_pkd));
1467 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1468 	if (cntxt_id >= sc->sge.neq)
1469 		panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1470 		    cntxt_id, sc->sge.neq - 1);
1471 	sc->sge.eqmap[cntxt_id] = eq;
1472 
1473 	return (rc);
1474 }
1475 
1476 #ifndef TCP_OFFLOAD_DISABLE
1477 static int
1478 ofld_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1479 {
1480 	int rc, cntxt_id;
1481 	struct fw_eq_ofld_cmd c;
1482 
1483 	bzero(&c, sizeof (c));
1484 
1485 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
1486 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
1487 	    V_FW_EQ_OFLD_CMD_VFN(0));
1488 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
1489 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
1490 	c.fetchszm_to_iqid =
1491 	    htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1492 	    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
1493 	    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
1494 	c.dcaen_to_eqsize =
1495 	    BE_32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1496 	    V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1497 	    V_FW_EQ_OFLD_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1498 	    V_FW_EQ_OFLD_CMD_EQSIZE(eq->qsize));
1499 	c.eqaddr = BE_64(eq->ba);
1500 
1501 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1502 	if (rc != 0) {
1503 		cxgb_printf(pi->dip, CE_WARN,
1504 		    "failed to create egress queue for TCP offload: %d", rc);
1505 		return (rc);
1506 	}
1507 	eq->flags |= EQ_ALLOCATED;
1508 
1509 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(BE_32(c.eqid_pkd));
1510 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1511 	if (cntxt_id >= sc->sge.neq)
1512 		panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1513 		    cntxt_id, sc->sge.neq - 1);
1514 	sc->sge.eqmap[cntxt_id] = eq;
1515 
1516 	return (rc);
1517 }
1518 #endif
1519 
1520 static int
1521 alloc_eq(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1522 {
1523 	int rc;
1524 	size_t len;
1525 
1526 	mutex_init(&eq->lock, NULL, MUTEX_DRIVER, DDI_INTR_PRI(sc->intr_pri));
1527 	eq->flags |= EQ_MTX;
1528 
1529 	len = eq->qsize * EQ_ESIZE;
1530 	rc = alloc_desc_ring(sc, len, DDI_DMA_WRITE, &eq->desc_dhdl,
1531 	    &eq->desc_ahdl, &eq->ba, (caddr_t *)&eq->desc);
1532 	if (rc != 0)
1533 		return (rc);
1534 
1535 	eq->cap = eq->qsize - SPG_SIZE / EQ_ESIZE;
1536 	eq->spg = (void *)&eq->desc[eq->cap];
1537 	eq->avail = eq->cap - 1;	/* one less to avoid cidx = pidx */
1538 	eq->pidx = eq->cidx = 0;
1539 
1540 	switch (eq->flags & EQ_TYPEMASK) {
1541 	case EQ_CTRL:
1542 		rc = ctrl_eq_alloc(sc, eq);
1543 		break;
1544 
1545 	case EQ_ETH:
1546 		rc = eth_eq_alloc(sc, pi, eq);
1547 		break;
1548 
1549 #ifndef TCP_OFFLOAD_DISABLE
1550 	case EQ_OFLD:
1551 		rc = ofld_eq_alloc(sc, pi, eq);
1552 		break;
1553 #endif
1554 
1555 	default:
1556 		panic("%s: invalid eq type %d.", __func__,
1557 		    eq->flags & EQ_TYPEMASK);
1558 	}
1559 	if (rc != 0) {
1560 		cxgb_printf(sc->dip, CE_WARN,
1561 		    "failed to allocate egress queue(%d): %d",
1562 		    eq->flags & EQ_TYPEMASK, rc);
1563 	}
1564 
1565 	return (rc);
1566 }
1567 
1568 static int
1569 free_eq(struct adapter *sc, struct sge_eq *eq)
1570 {
1571 	int rc;
1572 
1573 	if (eq->flags & EQ_ALLOCATED) {
1574 		switch (eq->flags & EQ_TYPEMASK) {
1575 		case EQ_CTRL:
1576 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
1577 			    eq->cntxt_id);
1578 			break;
1579 
1580 		case EQ_ETH:
1581 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
1582 			    eq->cntxt_id);
1583 			break;
1584 
1585 #ifndef TCP_OFFLOAD_DISABLE
1586 		case EQ_OFLD:
1587 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
1588 			    eq->cntxt_id);
1589 			break;
1590 #endif
1591 
1592 		default:
1593 			panic("%s: invalid eq type %d.", __func__,
1594 			    eq->flags & EQ_TYPEMASK);
1595 		}
1596 		if (rc != 0) {
1597 			cxgb_printf(sc->dip, CE_WARN,
1598 			    "failed to free egress queue (%d): %d",
1599 			    eq->flags & EQ_TYPEMASK, rc);
1600 			return (rc);
1601 		}
1602 		eq->flags &= ~EQ_ALLOCATED;
1603 	}
1604 
1605 	if (eq->desc != NULL) {
1606 		(void) free_desc_ring(&eq->desc_dhdl, &eq->desc_ahdl);
1607 		eq->desc = NULL;
1608 	}
1609 
1610 	if (eq->flags & EQ_MTX)
1611 		mutex_destroy(&eq->lock);
1612 
1613 	bzero(eq, sizeof (*eq));
1614 	return (0);
1615 }
1616 
1617 /* ARGSUSED */
1618 static int
1619 alloc_wrq(struct adapter *sc, struct port_info *pi, struct sge_wrq *wrq,
1620     int idx)
1621 {
1622 	int rc;
1623 
1624 	rc = alloc_eq(sc, pi, &wrq->eq);
1625 	if (rc != 0)
1626 		return (rc);
1627 
1628 	wrq->adapter = sc;
1629 	wrq->wr_list.head = NULL;
1630 	wrq->wr_list.tail = NULL;
1631 
1632 	/*
1633 	 * TODO: use idx to figure out what kind of wrq this is and install
1634 	 * useful kstats for it.
1635 	 */
1636 
1637 	return (rc);
1638 }
1639 
1640 static int
1641 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
1642 {
1643 	int rc;
1644 
1645 	rc = free_eq(sc, &wrq->eq);
1646 	if (rc != 0)
1647 		return (rc);
1648 
1649 	bzero(wrq, sizeof (*wrq));
1650 	return (0);
1651 }
1652 
1653 static int
1654 alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx)
1655 {
1656 	int rc, i;
1657 	struct adapter *sc = pi->adapter;
1658 	struct sge_eq *eq = &txq->eq;
1659 
1660 	rc = alloc_eq(sc, pi, eq);
1661 	if (rc != 0)
1662 		return (rc);
1663 
1664 	txq->port = pi;
1665 	txq->sdesc = kmem_zalloc(sizeof (struct tx_sdesc) * eq->cap, KM_SLEEP);
1666 	txq->txb_size = eq->qsize * tx_copy_threshold;
1667 	rc = alloc_tx_copybuffer(sc, txq->txb_size, &txq->txb_dhdl,
1668 	    &txq->txb_ahdl, &txq->txb_ba, &txq->txb_va);
1669 	if (rc == 0)
1670 		txq->txb_avail = txq->txb_size;
1671 	else
1672 		txq->txb_avail = txq->txb_size = 0;
1673 
1674 	/*
1675 	 * TODO: is this too low?  Worst case would need around 4 times qsize
1676 	 * (all tx descriptors filled to the brim with SGLs, with each entry in
1677 	 * the SGL coming from a distinct DMA handle).  Increase tx_dhdl_total
1678 	 * if you see too many dma_hdl_failed.
1679 	 */
1680 	txq->tx_dhdl_total = eq->qsize * 2;
1681 	txq->tx_dhdl = kmem_zalloc(sizeof (ddi_dma_handle_t) *
1682 	    txq->tx_dhdl_total, KM_SLEEP);
1683 	for (i = 0; i < txq->tx_dhdl_total; i++) {
1684 		rc = ddi_dma_alloc_handle(sc->dip, &sc->sge.dma_attr_tx,
1685 		    DDI_DMA_SLEEP, 0, &txq->tx_dhdl[i]);
1686 		if (rc != DDI_SUCCESS) {
1687 			cxgb_printf(sc->dip, CE_WARN,
1688 			    "%s: failed to allocate DMA handle (%d)",
1689 			    __func__, rc);
1690 			return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
1691 		}
1692 		txq->tx_dhdl_avail++;
1693 	}
1694 
1695 	txq->ksp = setup_txq_kstats(pi, txq, idx);
1696 
1697 	return (rc);
1698 }
1699 
1700 static int
1701 free_txq(struct port_info *pi, struct sge_txq *txq)
1702 {
1703 	int i;
1704 	struct adapter *sc = pi->adapter;
1705 	struct sge_eq *eq = &txq->eq;
1706 
1707 	if (txq->ksp != NULL) {
1708 		kstat_delete(txq->ksp);
1709 		txq->ksp = NULL;
1710 	}
1711 
1712 	if (txq->txb_va != NULL) {
1713 		(void) free_desc_ring(&txq->txb_dhdl, &txq->txb_ahdl);
1714 		txq->txb_va = NULL;
1715 	}
1716 
1717 	if (txq->sdesc != NULL) {
1718 		struct tx_sdesc *sd;
1719 		ddi_dma_handle_t hdl;
1720 
1721 		TXQ_LOCK(txq);
1722 		while (eq->cidx != eq->pidx) {
1723 			sd = &txq->sdesc[eq->cidx];
1724 
1725 			for (i = sd->hdls_used; i; i--) {
1726 				hdl = txq->tx_dhdl[txq->tx_dhdl_cidx];
1727 				(void) ddi_dma_unbind_handle(hdl);
1728 				if (++txq->tx_dhdl_cidx == txq->tx_dhdl_total)
1729 					txq->tx_dhdl_cidx = 0;
1730 			}
1731 
1732 			ASSERT(sd->m);
1733 			freemsgchain(sd->m);
1734 
1735 			eq->cidx += sd->desc_used;
1736 			if (eq->cidx >= eq->cap)
1737 				eq->cidx -= eq->cap;
1738 
1739 			txq->txb_avail += txq->txb_used;
1740 		}
1741 		ASSERT(txq->tx_dhdl_cidx == txq->tx_dhdl_pidx);
1742 		ASSERT(txq->txb_avail == txq->txb_size);
1743 		TXQ_UNLOCK(txq);
1744 
1745 		kmem_free(txq->sdesc, sizeof (struct tx_sdesc) * eq->cap);
1746 		txq->sdesc = NULL;
1747 	}
1748 
1749 	if (txq->tx_dhdl != NULL) {
1750 		for (i = 0; i < txq->tx_dhdl_total; i++) {
1751 			if (txq->tx_dhdl[i] != NULL)
1752 				ddi_dma_free_handle(&txq->tx_dhdl[i]);
1753 		}
1754 	}
1755 
1756 	(void) free_eq(sc, &txq->eq);
1757 
1758 	bzero(txq, sizeof (*txq));
1759 	return (0);
1760 }
1761 
1762 /*
1763  * Allocates a block of contiguous memory for DMA.  Can be used to allocate
1764  * memory for descriptor rings or for tx/rx copy buffers.
1765  *
1766  * Caller does not have to clean up anything if this function fails, it cleans
1767  * up after itself.
1768  *
1769  * Caller provides the following:
1770  * len		length of the block of memory to allocate.
1771  * flags	DDI_DMA_* flags to use (CONSISTENT/STREAMING, READ/WRITE/RDWR)
1772  * acc_attr	device access attributes for the allocation.
1773  * dma_attr	DMA attributes for the allocation
1774  *
1775  * If the function is successful it fills up this information:
1776  * dma_hdl	DMA handle for the allocated memory
1777  * acc_hdl	access handle for the allocated memory
1778  * ba		bus address of the allocated memory
1779  * va		KVA of the allocated memory.
1780  */
1781 static int
1782 alloc_dma_memory(struct adapter *sc, size_t len, int flags,
1783     ddi_device_acc_attr_t *acc_attr, ddi_dma_attr_t *dma_attr,
1784     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
1785     uint64_t *pba, caddr_t *pva)
1786 {
1787 	int rc;
1788 	ddi_dma_handle_t dhdl;
1789 	ddi_acc_handle_t ahdl;
1790 	ddi_dma_cookie_t cookie;
1791 	uint_t ccount;
1792 	caddr_t va;
1793 	size_t real_len;
1794 
1795 	*pva = NULL;
1796 
1797 	/*
1798 	 * DMA handle.
1799 	 */
1800 	rc = ddi_dma_alloc_handle(sc->dip, dma_attr, DDI_DMA_SLEEP, 0, &dhdl);
1801 	if (rc != DDI_SUCCESS) {
1802 		cxgb_printf(sc->dip, CE_WARN,
1803 		    "failed to allocate DMA handle: %d", rc);
1804 
1805 		return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
1806 	}
1807 
1808 	/*
1809 	 * Memory suitable for DMA.
1810 	 */
1811 	rc = ddi_dma_mem_alloc(dhdl, len, acc_attr,
1812 	    flags & DDI_DMA_CONSISTENT ? DDI_DMA_CONSISTENT : DDI_DMA_STREAMING,
1813 	    DDI_DMA_SLEEP, 0, &va, &real_len, &ahdl);
1814 	if (rc != DDI_SUCCESS) {
1815 		cxgb_printf(sc->dip, CE_WARN,
1816 		    "failed to allocate DMA memory: %d", rc);
1817 
1818 		ddi_dma_free_handle(&dhdl);
1819 		return (ENOMEM);
1820 	}
1821 
1822 	if (len != real_len) {
1823 		cxgb_printf(sc->dip, CE_WARN,
1824 		    "%s: len (%u) != real_len (%u)\n", len, real_len);
1825 	}
1826 
1827 	/*
1828 	 * DMA bindings.
1829 	 */
1830 	rc = ddi_dma_addr_bind_handle(dhdl, NULL, va, real_len, flags, NULL,
1831 	    NULL, &cookie, &ccount);
1832 	if (rc != DDI_DMA_MAPPED) {
1833 		cxgb_printf(sc->dip, CE_WARN,
1834 		    "failed to map DMA memory: %d", rc);
1835 
1836 		ddi_dma_mem_free(&ahdl);
1837 		ddi_dma_free_handle(&dhdl);
1838 		return (ENOMEM);
1839 	}
1840 	if (ccount != 1) {
1841 		cxgb_printf(sc->dip, CE_WARN,
1842 		    "unusable DMA mapping (%d segments)", ccount);
1843 		(void) free_desc_ring(&dhdl, &ahdl);
1844 	}
1845 
1846 	bzero(va, real_len);
1847 	*dma_hdl = dhdl;
1848 	*acc_hdl = ahdl;
1849 	*pba = cookie.dmac_laddress;
1850 	*pva = va;
1851 
1852 	return (0);
1853 }
1854 
1855 static int
1856 free_dma_memory(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl)
1857 {
1858 	(void) ddi_dma_unbind_handle(*dhdl);
1859 	ddi_dma_mem_free(ahdl);
1860 	ddi_dma_free_handle(dhdl);
1861 
1862 	return (0);
1863 }
1864 
1865 static int
1866 alloc_desc_ring(struct adapter *sc, size_t len, int rw,
1867     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
1868     uint64_t *pba, caddr_t *pva)
1869 {
1870 	ddi_device_acc_attr_t *acc_attr = &sc->sge.acc_attr_desc;
1871 	ddi_dma_attr_t *dma_attr = &sc->sge.dma_attr_desc;
1872 
1873 	return (alloc_dma_memory(sc, len, DDI_DMA_CONSISTENT | rw, acc_attr,
1874 	    dma_attr, dma_hdl, acc_hdl, pba, pva));
1875 }
1876 
1877 static int
1878 free_desc_ring(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl)
1879 {
1880 	return (free_dma_memory(dhdl, ahdl));
1881 }
1882 
1883 static int
1884 alloc_tx_copybuffer(struct adapter *sc, size_t len,
1885     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
1886     uint64_t *pba, caddr_t *pva)
1887 {
1888 	ddi_device_acc_attr_t *acc_attr = &sc->sge.acc_attr_tx;
1889 	ddi_dma_attr_t *dma_attr = &sc->sge.dma_attr_desc; /* NOT dma_attr_tx */
1890 
1891 	return (alloc_dma_memory(sc, len, DDI_DMA_STREAMING | DDI_DMA_WRITE,
1892 	    acc_attr, dma_attr, dma_hdl, acc_hdl, pba, pva));
1893 }
1894 
1895 static inline bool
1896 is_new_response(const struct sge_iq *iq, struct rsp_ctrl **ctrl)
1897 {
1898 	(void) ddi_dma_sync(iq->dhdl, (uintptr_t)iq->cdesc -
1899 	    (uintptr_t)iq->desc, iq->esize, DDI_DMA_SYNC_FORKERNEL);
1900 
1901 	*ctrl = (void *)((uintptr_t)iq->cdesc +
1902 	    (iq->esize - sizeof (struct rsp_ctrl)));
1903 
1904 	return ((((*ctrl)->u.type_gen >> S_RSPD_GEN) == iq->gen));
1905 }
1906 
1907 static inline void
1908 iq_next(struct sge_iq *iq)
1909 {
1910 	iq->cdesc = (void *) ((uintptr_t)iq->cdesc + iq->esize);
1911 	if (++iq->cidx == iq->qsize - 1) {
1912 		iq->cidx = 0;
1913 		iq->gen ^= 1;
1914 		iq->cdesc = iq->desc;
1915 	}
1916 }
1917 
1918 /*
1919  * Fill up the freelist by upto nbufs and maybe ring its doorbell.
1920  *
1921  * Returns non-zero to indicate that it should be added to the list of starving
1922  * freelists.
1923  */
1924 static int
1925 refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs)
1926 {
1927 	uint64_t *d = &fl->desc[fl->pidx];
1928 	struct fl_sdesc *sd = &fl->sdesc[fl->pidx];
1929 
1930 	FL_LOCK_ASSERT_OWNED(fl);
1931 	ASSERT(nbufs >= 0);
1932 
1933 	if (nbufs > fl->needed)
1934 		nbufs = fl->needed;
1935 
1936 	while (nbufs--) {
1937 		if (sd->rxb != NULL) {
1938 			if (sd->rxb->ref_cnt == 1) {
1939 				/*
1940 				 * Buffer is available for recycling.  Two ways
1941 				 * this can happen:
1942 				 *
1943 				 * a) All the packets DMA'd into it last time
1944 				 *    around were within the rx_copy_threshold
1945 				 *    and no part of the buffer was ever passed
1946 				 *    up (ref_cnt never went over 1).
1947 				 *
1948 				 * b) Packets DMA'd into the buffer were passed
1949 				 *    up but have all been freed by the upper
1950 				 *    layers by now (ref_cnt went over 1 but is
1951 				 *    now back to 1).
1952 				 *
1953 				 * Either way the bus address in the descriptor
1954 				 * ring is already valid.
1955 				 */
1956 				ASSERT(*d == cpu_to_be64(sd->rxb->ba));
1957 				d++;
1958 				goto recycled;
1959 			} else {
1960 				/*
1961 				 * Buffer still in use and we need a
1962 				 * replacement. But first release our reference
1963 				 * on the existing buffer.
1964 				 */
1965 				rxbuf_free(sd->rxb);
1966 			}
1967 		}
1968 
1969 		sd->rxb = rxbuf_alloc(sc->sge.rxbuf_cache, KM_NOSLEEP, 1);
1970 		if (sd->rxb == NULL)
1971 			break;
1972 		*d++ = cpu_to_be64(sd->rxb->ba);
1973 
1974 recycled:	fl->pending++;
1975 		sd++;
1976 		fl->needed--;
1977 		if (++fl->pidx == fl->cap) {
1978 			fl->pidx = 0;
1979 			sd = fl->sdesc;
1980 			d = fl->desc;
1981 		}
1982 	}
1983 
1984 	if (fl->pending >= 8)
1985 		ring_fl_db(sc, fl);
1986 
1987 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
1988 }
1989 
1990 #ifndef TAILQ_FOREACH_SAFE
1991 #define	TAILQ_FOREACH_SAFE(var, head, field, tvar)			\
1992 	for ((var) = TAILQ_FIRST((head));				\
1993 	    (var) && ((tvar) = TAILQ_NEXT((var), field), 1);		\
1994 	    (var) = (tvar))
1995 #endif
1996 
1997 /*
1998  * Attempt to refill all starving freelists.
1999  */
2000 static void
2001 refill_sfl(void *arg)
2002 {
2003 	struct adapter *sc = arg;
2004 	struct sge_fl *fl, *fl_temp;
2005 
2006 	mutex_enter(&sc->sfl_lock);
2007 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
2008 		FL_LOCK(fl);
2009 		(void) refill_fl(sc, fl, 64);
2010 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
2011 			TAILQ_REMOVE(&sc->sfl, fl, link);
2012 			fl->flags &= ~FL_STARVING;
2013 		}
2014 		FL_UNLOCK(fl);
2015 	}
2016 
2017 	if (!TAILQ_EMPTY(&sc->sfl) != 0)
2018 		sc->sfl_timer =  timeout(refill_sfl, sc, drv_usectohz(100000));
2019 	mutex_exit(&sc->sfl_lock);
2020 }
2021 
2022 static void
2023 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
2024 {
2025 	mutex_enter(&sc->sfl_lock);
2026 	FL_LOCK(fl);
2027 	if ((fl->flags & FL_DOOMED) == 0) {
2028 		if (TAILQ_EMPTY(&sc->sfl) != 0) {
2029 			sc->sfl_timer = timeout(refill_sfl, sc,
2030 			    drv_usectohz(100000));
2031 		}
2032 		fl->flags |= FL_STARVING;
2033 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
2034 	}
2035 	FL_UNLOCK(fl);
2036 	mutex_exit(&sc->sfl_lock);
2037 }
2038 
2039 static void
2040 free_fl_bufs(struct sge_fl *fl)
2041 {
2042 	struct fl_sdesc *sd;
2043 	unsigned int i;
2044 
2045 	FL_LOCK_ASSERT_OWNED(fl);
2046 
2047 	for (i = 0; i < fl->cap; i++) {
2048 		sd = &fl->sdesc[i];
2049 
2050 		if (sd->rxb != NULL) {
2051 			rxbuf_free(sd->rxb);
2052 			sd->rxb = NULL;
2053 		}
2054 	}
2055 }
2056 
2057 /*
2058  * Note that fl->cidx and fl->offset are left unchanged in case of failure.
2059  */
2060 static mblk_t *
2061 get_fl_payload(struct sge_fl *fl, uint32_t len_newbuf, int *fl_bufs_used)
2062 {
2063 	struct mblk_pair frame = {0};
2064 	struct rxbuf *rxb;
2065 	mblk_t *m = NULL;
2066 	uint_t nbuf = 0, len, copy, n;
2067 	uint32_t cidx, offset;
2068 
2069 	/*
2070 	 * The SGE won't pack a new frame into the current buffer if the entire
2071 	 * payload doesn't fit in the remaining space.  Move on to the next buf
2072 	 * in that case.
2073 	 */
2074 	if (fl->offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
2075 		fl->offset = 0;
2076 		if (++fl->cidx == fl->cap)
2077 			fl->cidx = 0;
2078 		nbuf++;
2079 	}
2080 	cidx = fl->cidx;
2081 	offset = fl->offset;
2082 
2083 	len = G_RSPD_LEN(len_newbuf);	/* pktshift + payload length */
2084 	copy = (len <= fl->copy_threshold);
2085 	if (copy != 0) {
2086 		frame.head = m = allocb(len, BPRI_HI);
2087 		if (m == NULL)
2088 			return (NULL);
2089 	}
2090 
2091 	while (len) {
2092 		rxb = fl->sdesc[cidx].rxb;
2093 		n = min(len, rxb->buf_size - offset);
2094 
2095 		(void) ddi_dma_sync(rxb->dhdl, offset, n,
2096 		    DDI_DMA_SYNC_FORKERNEL);
2097 
2098 		if (copy != 0)
2099 			bcopy(rxb->va + offset, m->b_wptr, n);
2100 		else {
2101 			m = desballoc((unsigned char *)rxb->va + offset, n,
2102 			    BPRI_HI, &rxb->freefunc);
2103 			if (m == NULL) {
2104 				freemsg(frame.head);
2105 				return (NULL);
2106 			}
2107 			atomic_inc_uint(&rxb->ref_cnt);
2108 			if (frame.head != NULL)
2109 				frame.tail->b_cont = m;
2110 			else
2111 				frame.head = m;
2112 			frame.tail = m;
2113 		}
2114 		m->b_wptr += n;
2115 		len -= n;
2116 		offset += roundup(n, FL_ALIGN);
2117 		ASSERT(offset <= rxb->buf_size);
2118 		if (offset == rxb->buf_size) {
2119 			offset = 0;
2120 			if (++cidx == fl->cap)
2121 				cidx = 0;
2122 			nbuf++;
2123 		}
2124 	}
2125 
2126 	fl->cidx = cidx;
2127 	fl->offset = offset;
2128 	(*fl_bufs_used) += nbuf;
2129 
2130 	ASSERT(frame.head != NULL);
2131 	return (frame.head);
2132 }
2133 
2134 /*
2135  * We'll do immediate data tx for non-LSO, but only when not coalescing.  We're
2136  * willing to use upto 2 hardware descriptors which means a maximum of 96 bytes
2137  * of immediate data.
2138  */
2139 #define	IMM_LEN ( \
2140 	2 * EQ_ESIZE \
2141 	- sizeof (struct fw_eth_tx_pkt_wr) \
2142 	- sizeof (struct cpl_tx_pkt_core))
2143 
2144 /*
2145  * Returns non-zero on failure, no need to cleanup anything in that case.
2146  *
2147  * Note 1: We always try to pull up the mblk if required and return E2BIG only
2148  * if this fails.
2149  *
2150  * Note 2: We'll also pullup incoming mblk if HW_LSO is set and the first mblk
2151  * does not have the TCP header in it.
2152  */
2153 static int
2154 get_frame_txinfo(struct sge_txq *txq, mblk_t **fp, struct txinfo *txinfo,
2155     int sgl_only)
2156 {
2157 	uint32_t flags = 0, len, n;
2158 	mblk_t *m = *fp;
2159 	int rc;
2160 
2161 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate txb and dma_hdls */
2162 
2163 	mac_hcksum_get(m, NULL, NULL, NULL, NULL, &flags);
2164 	txinfo->flags = flags;
2165 
2166 	mac_lso_get(m, &txinfo->mss, &flags);
2167 	txinfo->flags |= flags;
2168 
2169 	if (flags & HW_LSO)
2170 		sgl_only = 1;	/* Do not allow immediate data with LSO */
2171 
2172 start:	txinfo->nsegs = 0;
2173 	txinfo->hdls_used = 0;
2174 	txinfo->txb_used = 0;
2175 	txinfo->len = 0;
2176 
2177 	/* total length and a rough estimate of # of segments */
2178 	n = 0;
2179 	for (; m; m = m->b_cont) {
2180 		len = MBLKL(m);
2181 		n += (len / PAGE_SIZE) + 1;
2182 		txinfo->len += len;
2183 	}
2184 	m = *fp;
2185 
2186 	if (n >= TX_SGL_SEGS || (flags & HW_LSO && MBLKL(m) < 50)) {
2187 		txq->pullup_early++;
2188 		m = msgpullup(*fp, -1);
2189 		if (m == NULL) {
2190 			txq->pullup_failed++;
2191 			return (E2BIG);	/* (*fp) left as it was */
2192 		}
2193 		freemsg(*fp);
2194 		*fp = m;
2195 		mac_hcksum_set(m, NULL, NULL, NULL, NULL, txinfo->flags);
2196 	}
2197 
2198 	if (txinfo->len <= IMM_LEN && !sgl_only)
2199 		return (0);	/* nsegs = 0 tells caller to use imm. tx */
2200 
2201 	if (txinfo->len <= txq->copy_threshold &&
2202 	    copy_into_txb(txq, m, txinfo->len, txinfo) == 0)
2203 		goto done;
2204 
2205 	for (; m; m = m->b_cont) {
2206 
2207 		len = MBLKL(m);
2208 
2209 		/* Use tx copy buffer if this mblk is small enough */
2210 		if (len <= txq->copy_threshold &&
2211 		    copy_into_txb(txq, m, len, txinfo) == 0)
2212 			continue;
2213 
2214 		/* Add DMA bindings for this mblk to the SGL */
2215 		rc = add_mblk(txq, txinfo, m, len);
2216 
2217 		if (rc == E2BIG ||
2218 		    (txinfo->nsegs == TX_SGL_SEGS && m->b_cont)) {
2219 
2220 			txq->pullup_late++;
2221 			m = msgpullup(*fp, -1);
2222 			if (m != NULL) {
2223 				free_txinfo_resources(txq, txinfo);
2224 				freemsg(*fp);
2225 				*fp = m;
2226 				mac_hcksum_set(m, NULL, NULL, NULL, NULL,
2227 				    txinfo->flags);
2228 				goto start;
2229 			}
2230 
2231 			txq->pullup_failed++;
2232 			rc = E2BIG;
2233 		}
2234 
2235 		if (rc != 0) {
2236 			free_txinfo_resources(txq, txinfo);
2237 			return (rc);
2238 		}
2239 	}
2240 
2241 	ASSERT(txinfo->nsegs > 0 && txinfo->nsegs <= TX_SGL_SEGS);
2242 
2243 done:
2244 
2245 	/*
2246 	 * Store the # of flits required to hold this frame's SGL in nflits.  An
2247 	 * SGL has a (ULPTX header + len0, addr0) tuple optionally followed by
2248 	 * multiple (len0 + len1, addr0, addr1) tuples.  If addr1 is not used
2249 	 * then len1 must be set to 0.
2250 	 */
2251 	n = txinfo->nsegs - 1;
2252 	txinfo->nflits = (3 * n) / 2 + (n & 1) + 2;
2253 	if (n & 1)
2254 		txinfo->sgl.sge[n / 2].len[1] = cpu_to_be32(0);
2255 
2256 	txinfo->sgl.cmd_nsge = cpu_to_be32(V_ULPTX_CMD((u32)ULP_TX_SC_DSGL) |
2257 	    V_ULPTX_NSGE(txinfo->nsegs));
2258 
2259 	return (0);
2260 }
2261 
2262 static inline int
2263 fits_in_txb(struct sge_txq *txq, int len, int *waste)
2264 {
2265 	if (txq->txb_avail < len)
2266 		return (0);
2267 
2268 	if (txq->txb_next + len <= txq->txb_size) {
2269 		*waste = 0;
2270 		return (1);
2271 	}
2272 
2273 	*waste = txq->txb_size - txq->txb_next;
2274 
2275 	return (txq->txb_avail - *waste < len ? 0 : 1);
2276 }
2277 
2278 #define	TXB_CHUNK	64
2279 
2280 /*
2281  * Copies the specified # of bytes into txq's tx copy buffer and updates txinfo
2282  * and txq to indicate resources used.  Caller has to make sure that those many
2283  * bytes are available in the mblk chain (b_cont linked).
2284  */
2285 static inline int
2286 copy_into_txb(struct sge_txq *txq, mblk_t *m, int len, struct txinfo *txinfo)
2287 {
2288 	int waste, n;
2289 
2290 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate txb */
2291 
2292 	if (!fits_in_txb(txq, len, &waste)) {
2293 		txq->txb_full++;
2294 		return (ENOMEM);
2295 	}
2296 
2297 	if (waste != 0) {
2298 		ASSERT((waste & (TXB_CHUNK - 1)) == 0);
2299 		txinfo->txb_used += waste;
2300 		txq->txb_avail -= waste;
2301 		txq->txb_next = 0;
2302 	}
2303 
2304 	for (n = 0; n < len; m = m->b_cont) {
2305 		bcopy(m->b_rptr, txq->txb_va + txq->txb_next + n, MBLKL(m));
2306 		n += MBLKL(m);
2307 	}
2308 
2309 	add_seg(txinfo, txq->txb_ba + txq->txb_next, len);
2310 
2311 	n = roundup(len, TXB_CHUNK);
2312 	txinfo->txb_used += n;
2313 	txq->txb_avail -= n;
2314 	txq->txb_next += n;
2315 	ASSERT(txq->txb_next <= txq->txb_size);
2316 	if (txq->txb_next == txq->txb_size)
2317 		txq->txb_next = 0;
2318 
2319 	return (0);
2320 }
2321 
2322 static inline void
2323 add_seg(struct txinfo *txinfo, uint64_t ba, uint32_t len)
2324 {
2325 	ASSERT(txinfo->nsegs < TX_SGL_SEGS);	/* must have room */
2326 
2327 	if (txinfo->nsegs != 0) {
2328 		int idx = txinfo->nsegs - 1;
2329 		txinfo->sgl.sge[idx / 2].len[idx & 1] = cpu_to_be32(len);
2330 		txinfo->sgl.sge[idx / 2].addr[idx & 1] = cpu_to_be64(ba);
2331 	} else {
2332 		txinfo->sgl.len0 = cpu_to_be32(len);
2333 		txinfo->sgl.addr0 = cpu_to_be64(ba);
2334 	}
2335 	txinfo->nsegs++;
2336 }
2337 
2338 /*
2339  * This function cleans up any partially allocated resources when it fails so
2340  * there's nothing for the caller to clean up in that case.
2341  *
2342  * EIO indicates permanent failure.  Caller should drop the frame containing
2343  * this mblk and continue.
2344  *
2345  * E2BIG indicates that the SGL length for this mblk exceeds the hardware
2346  * limit.  Caller should pull up the frame before trying to send it out.
2347  * (This error means our pullup_early heuristic did not work for this frame)
2348  *
2349  * ENOMEM indicates a temporary shortage of resources (DMA handles, other DMA
2350  * resources, etc.).  Caller should suspend the tx queue and wait for reclaim to
2351  * free up resources.
2352  */
2353 static inline int
2354 add_mblk(struct sge_txq *txq, struct txinfo *txinfo, mblk_t *m, int len)
2355 {
2356 	ddi_dma_handle_t dhdl;
2357 	ddi_dma_cookie_t cookie;
2358 	uint_t ccount = 0;
2359 	int rc;
2360 
2361 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate dhdls */
2362 
2363 	if (txq->tx_dhdl_avail == 0) {
2364 		txq->dma_hdl_failed++;
2365 		return (ENOMEM);
2366 	}
2367 
2368 	dhdl = txq->tx_dhdl[txq->tx_dhdl_pidx];
2369 	rc = ddi_dma_addr_bind_handle(dhdl, NULL, (caddr_t)m->b_rptr, len,
2370 	    DDI_DMA_WRITE | DDI_DMA_STREAMING, DDI_DMA_DONTWAIT, NULL, &cookie,
2371 	    &ccount);
2372 	if (rc != DDI_DMA_MAPPED) {
2373 		txq->dma_map_failed++;
2374 
2375 		ASSERT(rc != DDI_DMA_INUSE && rc != DDI_DMA_PARTIAL_MAP);
2376 
2377 		return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EIO);
2378 	}
2379 
2380 	if (ccount + txinfo->nsegs > TX_SGL_SEGS) {
2381 		(void) ddi_dma_unbind_handle(dhdl);
2382 		return (E2BIG);
2383 	}
2384 
2385 	add_seg(txinfo, cookie.dmac_laddress, cookie.dmac_size);
2386 	while (--ccount) {
2387 		ddi_dma_nextcookie(dhdl, &cookie);
2388 		add_seg(txinfo, cookie.dmac_laddress, cookie.dmac_size);
2389 	}
2390 
2391 	if (++txq->tx_dhdl_pidx == txq->tx_dhdl_total)
2392 		txq->tx_dhdl_pidx = 0;
2393 	txq->tx_dhdl_avail--;
2394 	txinfo->hdls_used++;
2395 
2396 	return (0);
2397 }
2398 
2399 /*
2400  * Releases all the txq resources used up in the specified txinfo.
2401  */
2402 static void
2403 free_txinfo_resources(struct sge_txq *txq, struct txinfo *txinfo)
2404 {
2405 	int n;
2406 
2407 	TXQ_LOCK_ASSERT_OWNED(txq);	/* dhdls, txb */
2408 
2409 	n = txinfo->txb_used;
2410 	if (n > 0) {
2411 		txq->txb_avail += n;
2412 		if (n <= txq->txb_next)
2413 			txq->txb_next -= n;
2414 		else {
2415 			n -= txq->txb_next;
2416 			txq->txb_next = txq->txb_size - n;
2417 		}
2418 	}
2419 
2420 	for (n = txinfo->hdls_used; n > 0; n--) {
2421 		if (txq->tx_dhdl_pidx > 0)
2422 			txq->tx_dhdl_pidx--;
2423 		else
2424 			txq->tx_dhdl_pidx = txq->tx_dhdl_total - 1;
2425 		txq->tx_dhdl_avail++;
2426 		(void) ddi_dma_unbind_handle(txq->tx_dhdl[txq->tx_dhdl_pidx]);
2427 	}
2428 }
2429 
2430 /*
2431  * Returns 0 to indicate that m has been accepted into a coalesced tx work
2432  * request.  It has either been folded into txpkts or txpkts was flushed and m
2433  * has started a new coalesced work request (as the first frame in a fresh
2434  * txpkts).
2435  *
2436  * Returns non-zero to indicate a failure - caller is responsible for
2437  * transmitting m, if there was anything in txpkts it has been flushed.
2438  */
2439 static int
2440 add_to_txpkts(struct sge_txq *txq, struct txpkts *txpkts, mblk_t *m,
2441     struct txinfo *txinfo)
2442 {
2443 	struct sge_eq *eq = &txq->eq;
2444 	int can_coalesce;
2445 	struct tx_sdesc *txsd;
2446 	uint8_t flits;
2447 
2448 	TXQ_LOCK_ASSERT_OWNED(txq);
2449 
2450 	if (txpkts->npkt > 0) {
2451 		flits = TXPKTS_PKT_HDR + txinfo->nflits;
2452 		can_coalesce = (txinfo->flags & HW_LSO) == 0 &&
2453 		    txpkts->nflits + flits <= TX_WR_FLITS &&
2454 		    txpkts->nflits + flits <= eq->avail * 8 &&
2455 		    txpkts->plen + txinfo->len < 65536;
2456 
2457 		if (can_coalesce != 0) {
2458 			txpkts->tail->b_next = m;
2459 			txpkts->tail = m;
2460 			txpkts->npkt++;
2461 			txpkts->nflits += flits;
2462 			txpkts->plen += txinfo->len;
2463 
2464 			txsd = &txq->sdesc[eq->pidx];
2465 			txsd->txb_used += txinfo->txb_used;
2466 			txsd->hdls_used += txinfo->hdls_used;
2467 
2468 			return (0);
2469 		}
2470 
2471 		/*
2472 		 * Couldn't coalesce m into txpkts.  The first order of business
2473 		 * is to send txpkts on its way.  Then we'll revisit m.
2474 		 */
2475 		write_txpkts_wr(txq, txpkts);
2476 	}
2477 
2478 	/*
2479 	 * Check if we can start a new coalesced tx work request with m as
2480 	 * the first packet in it.
2481 	 */
2482 
2483 	ASSERT(txpkts->npkt == 0);
2484 	ASSERT(txinfo->len < 65536);
2485 
2486 	flits = TXPKTS_WR_HDR + txinfo->nflits;
2487 	can_coalesce = (txinfo->flags & HW_LSO) == 0 &&
2488 	    flits <= eq->avail * 8 && flits <= TX_WR_FLITS;
2489 
2490 	if (can_coalesce == 0)
2491 		return (EINVAL);
2492 
2493 	/*
2494 	 * Start a fresh coalesced tx WR with m as the first frame in it.
2495 	 */
2496 	txpkts->tail = m;
2497 	txpkts->npkt = 1;
2498 	txpkts->nflits = flits;
2499 	txpkts->flitp = &eq->desc[eq->pidx].flit[2];
2500 	txpkts->plen = txinfo->len;
2501 
2502 	txsd = &txq->sdesc[eq->pidx];
2503 	txsd->m = m;
2504 	txsd->txb_used = txinfo->txb_used;
2505 	txsd->hdls_used = txinfo->hdls_used;
2506 
2507 	return (0);
2508 }
2509 
2510 /*
2511  * Note that write_txpkts_wr can never run out of hardware descriptors (but
2512  * write_txpkt_wr can).  add_to_txpkts ensures that a frame is accepted for
2513  * coalescing only if sufficient hardware descriptors are available.
2514  */
2515 static void
2516 write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts)
2517 {
2518 	struct sge_eq *eq = &txq->eq;
2519 	struct fw_eth_tx_pkts_wr *wr;
2520 	struct tx_sdesc *txsd;
2521 	uint32_t ctrl;
2522 	uint16_t ndesc;
2523 
2524 	TXQ_LOCK_ASSERT_OWNED(txq);	/* pidx, avail */
2525 
2526 	ndesc = howmany(txpkts->nflits, 8);
2527 
2528 	wr = (void *)&eq->desc[eq->pidx];
2529 	wr->op_pkd = cpu_to_be32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR) |
2530 	    V_FW_WR_IMMDLEN(0)); /* immdlen does not matter in this WR */
2531 	ctrl = V_FW_WR_LEN16(howmany(txpkts->nflits, 2));
2532 	if (eq->avail == ndesc)
2533 		ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
2534 	wr->equiq_to_len16 = cpu_to_be32(ctrl);
2535 	wr->plen = cpu_to_be16(txpkts->plen);
2536 	wr->npkt = txpkts->npkt;
2537 	wr->r3 = wr->type = 0;
2538 
2539 	/* Everything else already written */
2540 
2541 	txsd = &txq->sdesc[eq->pidx];
2542 	txsd->desc_used = ndesc;
2543 
2544 	txq->txb_used += txsd->txb_used / TXB_CHUNK;
2545 	txq->hdl_used += txsd->hdls_used;
2546 
2547 	ASSERT(eq->avail >= ndesc);
2548 
2549 	eq->pending += ndesc;
2550 	eq->avail -= ndesc;
2551 	eq->pidx += ndesc;
2552 	if (eq->pidx >= eq->cap)
2553 		eq->pidx -= eq->cap;
2554 
2555 	txq->txpkts_pkts += txpkts->npkt;
2556 	txq->txpkts_wrs++;
2557 	txpkts->npkt = 0;	/* emptied */
2558 }
2559 
2560 static int
2561 write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, mblk_t *m,
2562     struct txinfo *txinfo)
2563 {
2564 	struct sge_eq *eq = &txq->eq;
2565 	struct fw_eth_tx_pkt_wr *wr;
2566 	struct cpl_tx_pkt_core *cpl;
2567 	uint32_t ctrl;	/* used in many unrelated places */
2568 	uint64_t ctrl1;
2569 	int nflits, ndesc;
2570 	struct tx_sdesc *txsd;
2571 	caddr_t dst;
2572 
2573 	TXQ_LOCK_ASSERT_OWNED(txq);	/* pidx, avail */
2574 
2575 	/*
2576 	 * Do we have enough flits to send this frame out?
2577 	 */
2578 	ctrl = sizeof (struct cpl_tx_pkt_core);
2579 	if (txinfo->flags & HW_LSO) {
2580 		nflits = TXPKT_LSO_WR_HDR;
2581 		ctrl += sizeof (struct cpl_tx_pkt_lso);
2582 	} else
2583 		nflits = TXPKT_WR_HDR;
2584 	if (txinfo->nsegs > 0)
2585 		nflits += txinfo->nflits;
2586 	else {
2587 		nflits += howmany(txinfo->len, 8);
2588 		ctrl += txinfo->len;
2589 	}
2590 	ndesc = howmany(nflits, 8);
2591 	if (ndesc > eq->avail)
2592 		return (ENOMEM);
2593 
2594 	/* Firmware work request header */
2595 	wr = (void *)&eq->desc[eq->pidx];
2596 	wr->op_immdlen = cpu_to_be32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
2597 	    V_FW_WR_IMMDLEN(ctrl));
2598 	ctrl = V_FW_WR_LEN16(howmany(nflits, 2));
2599 	if (eq->avail == ndesc)
2600 		ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
2601 	wr->equiq_to_len16 = cpu_to_be32(ctrl);
2602 	wr->r3 = 0;
2603 
2604 	if (txinfo->flags & HW_LSO) {
2605 		struct cpl_tx_pkt_lso *lso = (void *)(wr + 1);
2606 		char *p = (void *)m->b_rptr;
2607 		ctrl = V_LSO_OPCODE((u32)CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
2608 		    F_LSO_LAST_SLICE;
2609 
2610 		/* LINTED: E_BAD_PTR_CAST_ALIGN */
2611 		if (((struct ether_header *)p)->ether_type ==
2612 		    htons(ETHERTYPE_VLAN)) {
2613 			ctrl |= V_LSO_ETHHDR_LEN(1);
2614 			p += sizeof (struct ether_vlan_header);
2615 		} else
2616 			p += sizeof (struct ether_header);
2617 
2618 		/* LINTED: E_BAD_PTR_CAST_ALIGN for IPH_HDR_LENGTH() */
2619 		ctrl |= V_LSO_IPHDR_LEN(IPH_HDR_LENGTH(p) / 4);
2620 		/* LINTED: E_BAD_PTR_CAST_ALIGN for IPH_HDR_LENGTH() */
2621 		p += IPH_HDR_LENGTH(p);
2622 		ctrl |= V_LSO_TCPHDR_LEN(TCP_HDR_LENGTH((tcph_t *)p) / 4);
2623 
2624 		lso->lso_ctrl = cpu_to_be32(ctrl);
2625 		lso->ipid_ofst = cpu_to_be16(0);
2626 		lso->mss = cpu_to_be16(txinfo->mss);
2627 		lso->seqno_offset = cpu_to_be32(0);
2628 		lso->len = cpu_to_be32(txinfo->len);
2629 
2630 		cpl = (void *)(lso + 1);
2631 
2632 		txq->tso_wrs++;
2633 	} else
2634 		cpl = (void *)(wr + 1);
2635 
2636 	/* Checksum offload */
2637 	ctrl1 = 0;
2638 	if (!(txinfo->flags & HCK_IPV4_HDRCKSUM))
2639 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
2640 	if (!(txinfo->flags & HCK_FULLCKSUM))
2641 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
2642 	if (ctrl1 == 0)
2643 		txq->txcsum++;	/* some hardware assistance provided */
2644 
2645 	/* CPL header */
2646 	cpl->ctrl0 = cpu_to_be32(V_TXPKT_OPCODE(CPL_TX_PKT) |
2647 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
2648 	cpl->pack = 0;
2649 	cpl->len = cpu_to_be16(txinfo->len);
2650 	cpl->ctrl1 = cpu_to_be64(ctrl1);
2651 
2652 	/* Software descriptor */
2653 	txsd = &txq->sdesc[eq->pidx];
2654 	txsd->m = m;
2655 	txsd->txb_used = txinfo->txb_used;
2656 	txsd->hdls_used = txinfo->hdls_used;
2657 	/* LINTED: E_ASSIGN_NARROW_CONV */
2658 	txsd->desc_used = ndesc;
2659 
2660 	txq->txb_used += txinfo->txb_used / TXB_CHUNK;
2661 	txq->hdl_used += txinfo->hdls_used;
2662 
2663 	eq->pending += ndesc;
2664 	eq->avail -= ndesc;
2665 	eq->pidx += ndesc;
2666 	if (eq->pidx >= eq->cap)
2667 		eq->pidx -= eq->cap;
2668 
2669 	/* SGL */
2670 	dst = (void *)(cpl + 1);
2671 	if (txinfo->nsegs > 0) {
2672 		txq->sgl_wrs++;
2673 		copy_to_txd(eq, (void *)&txinfo->sgl, &dst, txinfo->nflits * 8);
2674 
2675 		/* Need to zero-pad to a 16 byte boundary if not on one */
2676 		if ((uintptr_t)dst & 0xf)
2677 			/* LINTED: E_BAD_PTR_CAST_ALIGN */
2678 			*(uint64_t *)dst = 0;
2679 
2680 	} else {
2681 		txq->imm_wrs++;
2682 #ifdef DEBUG
2683 		ctrl = txinfo->len;
2684 #endif
2685 		for (; m; m = m->b_cont) {
2686 			copy_to_txd(eq, (void *)m->b_rptr, &dst, MBLKL(m));
2687 #ifdef DEBUG
2688 			ctrl -= MBLKL(m);
2689 #endif
2690 		}
2691 		ASSERT(ctrl == 0);
2692 	}
2693 
2694 	txq->txpkt_wrs++;
2695 	return (0);
2696 }
2697 
2698 static inline void
2699 write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
2700     struct txpkts *txpkts, struct txinfo *txinfo)
2701 {
2702 	struct ulp_txpkt *ulpmc;
2703 	struct ulptx_idata *ulpsc;
2704 	struct cpl_tx_pkt_core *cpl;
2705 	uintptr_t flitp, start, end;
2706 	uint64_t ctrl;
2707 	caddr_t dst;
2708 
2709 	ASSERT(txpkts->npkt > 0);
2710 
2711 	start = (uintptr_t)txq->eq.desc;
2712 	end = (uintptr_t)txq->eq.spg;
2713 
2714 	/* Checksum offload */
2715 	ctrl = 0;
2716 	if (!(txinfo->flags & HCK_IPV4_HDRCKSUM))
2717 		ctrl |= F_TXPKT_IPCSUM_DIS;
2718 	if (!(txinfo->flags & HCK_FULLCKSUM))
2719 		ctrl |= F_TXPKT_L4CSUM_DIS;
2720 	if (ctrl == 0)
2721 		txq->txcsum++;	/* some hardware assistance provided */
2722 
2723 	/*
2724 	 * The previous packet's SGL must have ended at a 16 byte boundary (this
2725 	 * is required by the firmware/hardware).  It follows that flitp cannot
2726 	 * wrap around between the ULPTX master command and ULPTX subcommand (8
2727 	 * bytes each), and that it can not wrap around in the middle of the
2728 	 * cpl_tx_pkt_core either.
2729 	 */
2730 	flitp = (uintptr_t)txpkts->flitp;
2731 	ASSERT((flitp & 0xf) == 0);
2732 
2733 	/* ULP master command */
2734 	ulpmc = (void *)flitp;
2735 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
2736 	ulpmc->len = htonl(howmany(sizeof (*ulpmc) + sizeof (*ulpsc) +
2737 	    sizeof (*cpl) + 8 * txinfo->nflits, 16));
2738 
2739 	/* ULP subcommand */
2740 	ulpsc = (void *)(ulpmc + 1);
2741 	ulpsc->cmd_more = cpu_to_be32(V_ULPTX_CMD((u32)ULP_TX_SC_IMM) |
2742 	    F_ULP_TX_SC_MORE);
2743 	ulpsc->len = cpu_to_be32(sizeof (struct cpl_tx_pkt_core));
2744 
2745 	flitp += sizeof (*ulpmc) + sizeof (*ulpsc);
2746 	if (flitp == end)
2747 		flitp = start;
2748 
2749 	/* CPL_TX_PKT */
2750 	cpl = (void *)flitp;
2751 	cpl->ctrl0 = cpu_to_be32(V_TXPKT_OPCODE(CPL_TX_PKT) |
2752 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
2753 	cpl->pack = 0;
2754 	cpl->len = cpu_to_be16(txinfo->len);
2755 	cpl->ctrl1 = cpu_to_be64(ctrl);
2756 
2757 	flitp += sizeof (*cpl);
2758 	if (flitp == end)
2759 		flitp = start;
2760 
2761 	/* SGL for this frame */
2762 	dst = (caddr_t)flitp;
2763 	copy_to_txd(&txq->eq, (void *)&txinfo->sgl, &dst, txinfo->nflits * 8);
2764 	flitp = (uintptr_t)dst;
2765 
2766 	/* Zero pad and advance to a 16 byte boundary if not already at one. */
2767 	if (flitp & 0xf) {
2768 
2769 		/* no matter what, flitp should be on an 8 byte boundary */
2770 		ASSERT((flitp & 0x7) == 0);
2771 
2772 		*(uint64_t *)flitp = 0;
2773 		flitp += sizeof (uint64_t);
2774 		txpkts->nflits++;
2775 	}
2776 
2777 	if (flitp == end)
2778 		flitp = start;
2779 
2780 	txpkts->flitp = (void *)flitp;
2781 }
2782 
2783 static inline void
2784 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
2785 {
2786 	if ((uintptr_t)(*to) + len <= (uintptr_t)eq->spg) {
2787 		bcopy(from, *to, len);
2788 		(*to) += len;
2789 	} else {
2790 		int portion = (uintptr_t)eq->spg - (uintptr_t)(*to);
2791 
2792 		bcopy(from, *to, portion);
2793 		from += portion;
2794 		portion = len - portion;	/* remaining */
2795 		bcopy(from, (void *)eq->desc, portion);
2796 		(*to) = (caddr_t)eq->desc + portion;
2797 	}
2798 }
2799 
2800 static inline void
2801 ring_tx_db(struct adapter *sc, struct sge_eq *eq)
2802 {
2803 	if (eq->pending > eq->pidx) {
2804 		int offset = eq->cap - (eq->pending - eq->pidx);
2805 
2806 		/* pidx has wrapped around since last doorbell */
2807 
2808 		(void) ddi_dma_sync(eq->desc_dhdl,
2809 		    offset * sizeof (struct tx_desc), 0,
2810 		    DDI_DMA_SYNC_FORDEV);
2811 		(void) ddi_dma_sync(eq->desc_dhdl,
2812 		    0, eq->pidx * sizeof (struct tx_desc),
2813 		    DDI_DMA_SYNC_FORDEV);
2814 	} else if (eq->pending > 0) {
2815 		(void) ddi_dma_sync(eq->desc_dhdl,
2816 		    (eq->pidx - eq->pending) * sizeof (struct tx_desc),
2817 		    eq->pending * sizeof (struct tx_desc),
2818 		    DDI_DMA_SYNC_FORDEV);
2819 	}
2820 
2821 	membar_producer();
2822 
2823 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
2824 	    V_QID(eq->cntxt_id) | V_PIDX(eq->pending));
2825 
2826 	eq->pending = 0;
2827 }
2828 
2829 static int
2830 reclaim_tx_descs(struct sge_txq *txq, int howmany)
2831 {
2832 	struct tx_sdesc *txsd;
2833 	uint_t cidx, can_reclaim, reclaimed, txb_freed, hdls_freed;
2834 	struct sge_eq *eq = &txq->eq;
2835 
2836 	EQ_LOCK_ASSERT_OWNED(eq);
2837 
2838 	cidx = eq->spg->cidx;	/* stable snapshot */
2839 	cidx = be16_to_cpu(cidx);
2840 
2841 	if (cidx >= eq->cidx)
2842 		can_reclaim = cidx - eq->cidx;
2843 	else
2844 		can_reclaim = cidx + eq->cap - eq->cidx;
2845 
2846 	if (can_reclaim == 0)
2847 		return (0);
2848 
2849 	txb_freed = hdls_freed = reclaimed = 0;
2850 	do {
2851 		int ndesc;
2852 
2853 		txsd = &txq->sdesc[eq->cidx];
2854 		ndesc = txsd->desc_used;
2855 
2856 		/* Firmware doesn't return "partial" credits. */
2857 		ASSERT(can_reclaim >= ndesc);
2858 
2859 		/*
2860 		 * We always keep mblk around, even for immediate data.  If mblk
2861 		 * is NULL, this has to be the software descriptor for a credit
2862 		 * flush work request.
2863 		 */
2864 		if (txsd->m != NULL)
2865 			freemsgchain(txsd->m);
2866 #ifdef DEBUG
2867 		else {
2868 			ASSERT(txsd->txb_used == 0);
2869 			ASSERT(txsd->hdls_used == 0);
2870 			ASSERT(ndesc == 1);
2871 		}
2872 #endif
2873 
2874 		txb_freed += txsd->txb_used;
2875 		hdls_freed += txsd->hdls_used;
2876 		reclaimed += ndesc;
2877 
2878 		eq->cidx += ndesc;
2879 		if (eq->cidx >= eq->cap)
2880 			eq->cidx -= eq->cap;
2881 
2882 		can_reclaim -= ndesc;
2883 
2884 	} while (can_reclaim && reclaimed < howmany);
2885 
2886 	eq->avail += reclaimed;
2887 	ASSERT(eq->avail < eq->cap);	/* avail tops out at (cap - 1) */
2888 
2889 	txq->txb_avail += txb_freed;
2890 
2891 	txq->tx_dhdl_avail += hdls_freed;
2892 	ASSERT(txq->tx_dhdl_avail <= txq->tx_dhdl_total);
2893 	for (; hdls_freed; hdls_freed--) {
2894 		(void) ddi_dma_unbind_handle(txq->tx_dhdl[txq->tx_dhdl_cidx]);
2895 		if (++txq->tx_dhdl_cidx == txq->tx_dhdl_total)
2896 			txq->tx_dhdl_cidx = 0;
2897 	}
2898 
2899 	return (reclaimed);
2900 }
2901 
2902 static void
2903 write_txqflush_wr(struct sge_txq *txq)
2904 {
2905 	struct sge_eq *eq = &txq->eq;
2906 	struct fw_eq_flush_wr *wr;
2907 	struct tx_sdesc *txsd;
2908 
2909 	EQ_LOCK_ASSERT_OWNED(eq);
2910 	ASSERT(eq->avail > 0);
2911 
2912 	wr = (void *)&eq->desc[eq->pidx];
2913 	bzero(wr, sizeof (*wr));
2914 	wr->opcode = FW_EQ_FLUSH_WR;
2915 	wr->equiq_to_len16 = cpu_to_be32(V_FW_WR_LEN16(sizeof (*wr) / 16) |
2916 	    F_FW_WR_EQUEQ | F_FW_WR_EQUIQ);
2917 
2918 	txsd = &txq->sdesc[eq->pidx];
2919 	txsd->m = NULL;
2920 	txsd->txb_used = 0;
2921 	txsd->hdls_used = 0;
2922 	txsd->desc_used = 1;
2923 
2924 	eq->pending++;
2925 	eq->avail--;
2926 	if (++eq->pidx == eq->cap)
2927 		eq->pidx = 0;
2928 }
2929 
2930 static int
2931 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, mblk_t *m)
2932 {
2933 	struct sge_rxq *rxq = (void *)iq;
2934 	struct mblk_pair chain = {0};
2935 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
2936 
2937 	iq->intr_next = iq->intr_params;
2938 
2939 	m->b_rptr += FL_PKTSHIFT;
2940 
2941 	/* TODO: what about cpl->ip_frag? */
2942 	if (cpl->csum_calc && !cpl->err_vec && !cpl->ip_frag) {
2943 		mac_hcksum_set(m, 0, 0, 0, 0xffff,
2944 		    HCK_FULLCKSUM_OK | HCK_FULLCKSUM |
2945 		    HCK_IPV4_HDRCKSUM_OK);
2946 		rxq->rxcsum++;
2947 	}
2948 
2949 	/* Add to the chain that we'll send up */
2950 	if (chain.head != NULL)
2951 		chain.tail->b_next = m;
2952 	else
2953 		chain.head = m;
2954 	chain.tail = m;
2955 
2956 	t4_mac_rx(rxq->port, rxq, chain.head);
2957 
2958 	return (0);
2959 }
2960 
2961 #define	FL_HW_IDX(idx)	((idx) >> 3)
2962 
2963 static inline void
2964 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
2965 {
2966 	int desc_start, desc_last, ndesc;
2967 
2968 	ndesc = FL_HW_IDX(fl->pending);
2969 
2970 	/* Hold back one credit if pidx = cidx */
2971 	if (FL_HW_IDX(fl->pidx) == FL_HW_IDX(fl->cidx))
2972 		ndesc--;
2973 
2974 	/*
2975 	 * There are chances of ndesc modified above (to avoid pidx = cidx).
2976 	 * If there is nothing to post, return.
2977 	 */
2978 	if (ndesc <= 0)
2979 		return;
2980 
2981 	desc_last = FL_HW_IDX(fl->pidx);
2982 
2983 	if (fl->pidx < fl->pending) {
2984 		/* There was a wrap */
2985 		desc_start = FL_HW_IDX(fl->pidx + fl->cap - fl->pending);
2986 
2987 		/* From desc_start to the end of list */
2988 		(void) ddi_dma_sync(fl->dhdl, desc_start * RX_FL_ESIZE, 0,
2989 		    DDI_DMA_SYNC_FORDEV);
2990 
2991 		/* From start of list to the desc_last */
2992 		if (desc_last != 0)
2993 			(void) ddi_dma_sync(fl->dhdl, 0, desc_last *
2994 			    RX_FL_ESIZE, DDI_DMA_SYNC_FORDEV);
2995 	} else {
2996 		/* There was no wrap, sync from start_desc to last_desc */
2997 		desc_start = FL_HW_IDX(fl->pidx - fl->pending);
2998 		(void) ddi_dma_sync(fl->dhdl, desc_start * RX_FL_ESIZE,
2999 		    ndesc * RX_FL_ESIZE, DDI_DMA_SYNC_FORDEV);
3000 	}
3001 
3002 	membar_producer();
3003 
3004 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), F_DBPRIO |
3005 	    V_QID(fl->cntxt_id) | V_PIDX(ndesc));
3006 
3007 	/*
3008 	 * Update pending count:
3009 	 * Deduct the number of descriptors posted
3010 	 */
3011 	fl->pending -= ndesc * 8;
3012 }
3013 
3014 /* ARGSUSED */
3015 static int
3016 handle_fw_rpl(struct sge_iq *iq, const struct rss_header *rss, mblk_t *m)
3017 {
3018 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
3019 
3020 	ASSERT(m == NULL);
3021 
3022 	if (cpl->type == FW6_TYPE_CMD_RPL)
3023 		(void) t4_handle_fw_rpl(iq->adapter, cpl->data);
3024 
3025 	return (0);
3026 }
3027 
3028 int
3029 t4_alloc_tx_maps(struct adapter *sc, struct tx_maps *txmaps, int count,
3030     int flags)
3031 {
3032 	int i, rc;
3033 
3034 	txmaps->map_total =  count;
3035 	txmaps->map_avail = txmaps->map_cidx = txmaps->map_pidx = 0;
3036 
3037 	txmaps->map =  kmem_zalloc(sizeof (ddi_dma_handle_t) *
3038 	    txmaps->map_total, flags);
3039 
3040 	for (i = 0; i < count; i++) {
3041 		rc = ddi_dma_alloc_handle(sc->dip, &sc->sge.dma_attr_tx,
3042 		    DDI_DMA_SLEEP, 0, &txmaps->map[i]);
3043 		if (rc != DDI_SUCCESS) {
3044 			cxgb_printf(sc->dip, CE_WARN,
3045 			    "%s: failed to allocate DMA handle (%d)",
3046 			    __func__, rc);
3047 			return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
3048 		}
3049 		txmaps->map_avail++;
3050 	}
3051 
3052 	return (0);
3053 }
3054 
3055 #define	KS_UINIT(x)	kstat_named_init(&kstatp->x, #x, KSTAT_DATA_ULONG)
3056 #define	KS_CINIT(x)	kstat_named_init(&kstatp->x, #x, KSTAT_DATA_CHAR)
3057 #define	KS_U_SET(x, y)	kstatp->x.value.ul = (y)
3058 #define	KS_U_FROM(x, y)	kstatp->x.value.ul = (y)->x
3059 #define	KS_C_SET(x, ...)	\
3060 			(void) snprintf(kstatp->x.value.c, 16,  __VA_ARGS__)
3061 
3062 /*
3063  * cxgbe:X:config
3064  */
3065 struct cxgbe_port_config_kstats {
3066 	kstat_named_t idx;
3067 	kstat_named_t nrxq;
3068 	kstat_named_t ntxq;
3069 	kstat_named_t first_rxq;
3070 	kstat_named_t first_txq;
3071 	kstat_named_t controller;
3072 	kstat_named_t factory_mac_address;
3073 };
3074 
3075 /*
3076  * cxgbe:X:info
3077  */
3078 struct cxgbe_port_info_kstats {
3079 	kstat_named_t transceiver;
3080 	kstat_named_t rx_ovflow0;
3081 	kstat_named_t rx_ovflow1;
3082 	kstat_named_t rx_ovflow2;
3083 	kstat_named_t rx_ovflow3;
3084 	kstat_named_t rx_trunc0;
3085 	kstat_named_t rx_trunc1;
3086 	kstat_named_t rx_trunc2;
3087 	kstat_named_t rx_trunc3;
3088 	kstat_named_t tx_pause;
3089 	kstat_named_t rx_pause;
3090 };
3091 
3092 static kstat_t *
3093 setup_port_config_kstats(struct port_info *pi)
3094 {
3095 	kstat_t *ksp;
3096 	struct cxgbe_port_config_kstats *kstatp;
3097 	int ndata;
3098 	dev_info_t *pdip = ddi_get_parent(pi->dip);
3099 	uint8_t *ma = &pi->hw_addr[0];
3100 
3101 	ndata = sizeof (struct cxgbe_port_config_kstats) /
3102 	    sizeof (kstat_named_t);
3103 
3104 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), "config",
3105 	    "net", KSTAT_TYPE_NAMED, ndata, 0);
3106 	if (ksp == NULL) {
3107 		cxgb_printf(pi->dip, CE_WARN, "failed to initialize kstats.");
3108 		return (NULL);
3109 	}
3110 
3111 	kstatp = (struct cxgbe_port_config_kstats *)ksp->ks_data;
3112 
3113 	KS_UINIT(idx);
3114 	KS_UINIT(nrxq);
3115 	KS_UINIT(ntxq);
3116 	KS_UINIT(first_rxq);
3117 	KS_UINIT(first_txq);
3118 	KS_CINIT(controller);
3119 	KS_CINIT(factory_mac_address);
3120 
3121 	KS_U_SET(idx, pi->port_id);
3122 	KS_U_SET(nrxq, pi->nrxq);
3123 	KS_U_SET(ntxq, pi->ntxq);
3124 	KS_U_SET(first_rxq, pi->first_rxq);
3125 	KS_U_SET(first_txq, pi->first_txq);
3126 	KS_C_SET(controller, "%s%d", ddi_driver_name(pdip),
3127 	    ddi_get_instance(pdip));
3128 	KS_C_SET(factory_mac_address, "%02X%02X%02X%02X%02X%02X",
3129 	    ma[0], ma[1], ma[2], ma[3], ma[4], ma[5]);
3130 
3131 	/* Do NOT set ksp->ks_update.  These kstats do not change. */
3132 
3133 	/* Install the kstat */
3134 	ksp->ks_private = (void *)pi;
3135 	kstat_install(ksp);
3136 
3137 	return (ksp);
3138 }
3139 
3140 static kstat_t *
3141 setup_port_info_kstats(struct port_info *pi)
3142 {
3143 	kstat_t *ksp;
3144 	struct cxgbe_port_info_kstats *kstatp;
3145 	int ndata;
3146 
3147 	ndata = sizeof (struct cxgbe_port_info_kstats) / sizeof (kstat_named_t);
3148 
3149 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), "info",
3150 	    "net", KSTAT_TYPE_NAMED, ndata, 0);
3151 	if (ksp == NULL) {
3152 		cxgb_printf(pi->dip, CE_WARN, "failed to initialize kstats.");
3153 		return (NULL);
3154 	}
3155 
3156 	kstatp = (struct cxgbe_port_info_kstats *)ksp->ks_data;
3157 
3158 	KS_CINIT(transceiver);
3159 	KS_UINIT(rx_ovflow0);
3160 	KS_UINIT(rx_ovflow1);
3161 	KS_UINIT(rx_ovflow2);
3162 	KS_UINIT(rx_ovflow3);
3163 	KS_UINIT(rx_trunc0);
3164 	KS_UINIT(rx_trunc1);
3165 	KS_UINIT(rx_trunc2);
3166 	KS_UINIT(rx_trunc3);
3167 	KS_UINIT(tx_pause);
3168 	KS_UINIT(rx_pause);
3169 
3170 	/* Install the kstat */
3171 	ksp->ks_update = update_port_info_kstats;
3172 	ksp->ks_private = (void *)pi;
3173 	kstat_install(ksp);
3174 
3175 	return (ksp);
3176 }
3177 
3178 static int
3179 update_port_info_kstats(kstat_t *ksp, int rw)
3180 {
3181 	struct cxgbe_port_info_kstats *kstatp =
3182 	    (struct cxgbe_port_info_kstats *)ksp->ks_data;
3183 	struct port_info *pi = ksp->ks_private;
3184 	static const char *mod_str[] = { NULL, "LR", "SR", "ER", "TWINAX",
3185 	    "active TWINAX", "LRM" };
3186 	uint32_t bgmap;
3187 
3188 	if (rw == KSTAT_WRITE)
3189 		return (0);
3190 
3191 	if (pi->mod_type == FW_PORT_MOD_TYPE_NONE)
3192 		KS_C_SET(transceiver, "unplugged");
3193 	else if (pi->mod_type == FW_PORT_MOD_TYPE_UNKNOWN)
3194 		KS_C_SET(transceiver, "unknown");
3195 	else if (pi->mod_type == FW_PORT_MOD_TYPE_NOTSUPPORTED)
3196 		KS_C_SET(transceiver, "unsupported");
3197 	else if (pi->mod_type > 0 && pi->mod_type < ARRAY_SIZE(mod_str))
3198 		KS_C_SET(transceiver, "%s", mod_str[pi->mod_type]);
3199 	else
3200 		KS_C_SET(transceiver, "type %d", pi->mod_type);
3201 
3202 #define	GET_STAT(name) t4_read_reg64(pi->adapter, \
3203 	    PORT_REG(pi->port_id, A_MPS_PORT_STAT_##name##_L))
3204 #define	GET_STAT_COM(name) t4_read_reg64(pi->adapter, \
3205 	    A_MPS_STAT_##name##_L)
3206 
3207 	bgmap = G_NUMPORTS(t4_read_reg(pi->adapter, A_MPS_CMN_CTL));
3208 	if (bgmap == 0)
3209 		bgmap = (pi->port_id == 0) ? 0xf : 0;
3210 	else if (bgmap == 1)
3211 		bgmap = (pi->port_id < 2) ? (3 << (2 * pi->port_id)) : 0;
3212 	else
3213 		bgmap = 1;
3214 
3215 	KS_U_SET(rx_ovflow0, (bgmap & 1) ?
3216 	    GET_STAT_COM(RX_BG_0_MAC_DROP_FRAME) : 0);
3217 	KS_U_SET(rx_ovflow1, (bgmap & 2) ?
3218 	    GET_STAT_COM(RX_BG_1_MAC_DROP_FRAME) : 0);
3219 	KS_U_SET(rx_ovflow2, (bgmap & 4) ?
3220 	    GET_STAT_COM(RX_BG_2_MAC_DROP_FRAME) : 0);
3221 	KS_U_SET(rx_ovflow3, (bgmap & 8) ?
3222 	    GET_STAT_COM(RX_BG_3_MAC_DROP_FRAME) : 0);
3223 	KS_U_SET(rx_trunc0,  (bgmap & 1) ?
3224 	    GET_STAT_COM(RX_BG_0_MAC_TRUNC_FRAME) : 0);
3225 	KS_U_SET(rx_trunc1,  (bgmap & 2) ?
3226 	    GET_STAT_COM(RX_BG_1_MAC_TRUNC_FRAME) : 0);
3227 	KS_U_SET(rx_trunc2,  (bgmap & 4) ?
3228 	    GET_STAT_COM(RX_BG_2_MAC_TRUNC_FRAME) : 0);
3229 	KS_U_SET(rx_trunc3,  (bgmap & 8) ?
3230 	    GET_STAT_COM(RX_BG_3_MAC_TRUNC_FRAME) : 0);
3231 
3232 	KS_U_SET(tx_pause, GET_STAT(TX_PORT_PAUSE));
3233 	KS_U_SET(rx_pause, GET_STAT(RX_PORT_PAUSE));
3234 
3235 	return (0);
3236 
3237 }
3238 
3239 /*
3240  * cxgbe:X:rxqY
3241  */
3242 struct rxq_kstats {
3243 	kstat_named_t rxcsum;
3244 	kstat_named_t nomem;
3245 };
3246 
3247 static kstat_t *
3248 setup_rxq_kstats(struct port_info *pi, struct sge_rxq *rxq, int idx)
3249 {
3250 	struct kstat *ksp;
3251 	struct rxq_kstats *kstatp;
3252 	int ndata;
3253 	char str[16];
3254 
3255 	ndata = sizeof (struct rxq_kstats) / sizeof (kstat_named_t);
3256 	(void) snprintf(str, sizeof (str), "rxq%u", idx);
3257 
3258 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), str, "rxq",
3259 	    KSTAT_TYPE_NAMED, ndata, 0);
3260 	if (ksp == NULL) {
3261 		cxgb_printf(pi->dip, CE_WARN,
3262 		    "%s: failed to initialize rxq kstats for queue %d.",
3263 		    __func__, idx);
3264 		return (NULL);
3265 	}
3266 
3267 	kstatp = (struct rxq_kstats *)ksp->ks_data;
3268 
3269 	KS_UINIT(rxcsum);
3270 	KS_UINIT(nomem);
3271 
3272 	ksp->ks_update = update_rxq_kstats;
3273 	ksp->ks_private = (void *)rxq;
3274 	kstat_install(ksp);
3275 
3276 	return (ksp);
3277 }
3278 
3279 static int
3280 update_rxq_kstats(kstat_t *ksp, int rw)
3281 {
3282 	struct rxq_kstats *kstatp = (struct rxq_kstats *)ksp->ks_data;
3283 	struct sge_rxq *rxq = ksp->ks_private;
3284 
3285 	if (rw == KSTAT_WRITE)
3286 		return (0);
3287 
3288 	KS_U_FROM(rxcsum, rxq);
3289 	KS_U_FROM(nomem, rxq);
3290 
3291 	return (0);
3292 }
3293 
3294 /*
3295  * cxgbe:X:txqY
3296  */
3297 struct txq_kstats {
3298 	kstat_named_t txcsum;
3299 	kstat_named_t tso_wrs;
3300 	kstat_named_t imm_wrs;
3301 	kstat_named_t sgl_wrs;
3302 	kstat_named_t txpkt_wrs;
3303 	kstat_named_t txpkts_wrs;
3304 	kstat_named_t txpkts_pkts;
3305 	kstat_named_t txb_used;
3306 	kstat_named_t hdl_used;
3307 	kstat_named_t txb_full;
3308 	kstat_named_t dma_hdl_failed;
3309 	kstat_named_t dma_map_failed;
3310 	kstat_named_t qfull;
3311 	kstat_named_t qflush;
3312 	kstat_named_t pullup_early;
3313 	kstat_named_t pullup_late;
3314 	kstat_named_t pullup_failed;
3315 };
3316 
3317 static kstat_t *
3318 setup_txq_kstats(struct port_info *pi, struct sge_txq *txq, int idx)
3319 {
3320 	struct kstat *ksp;
3321 	struct txq_kstats *kstatp;
3322 	int ndata;
3323 	char str[16];
3324 
3325 	ndata = sizeof (struct txq_kstats) / sizeof (kstat_named_t);
3326 	(void) snprintf(str, sizeof (str), "txq%u", idx);
3327 
3328 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), str, "txq",
3329 	    KSTAT_TYPE_NAMED, ndata, 0);
3330 	if (ksp == NULL) {
3331 		cxgb_printf(pi->dip, CE_WARN,
3332 		    "%s: failed to initialize txq kstats for queue %d.",
3333 		    __func__, idx);
3334 		return (NULL);
3335 	}
3336 
3337 	kstatp = (struct txq_kstats *)ksp->ks_data;
3338 
3339 	KS_UINIT(txcsum);
3340 	KS_UINIT(tso_wrs);
3341 	KS_UINIT(imm_wrs);
3342 	KS_UINIT(sgl_wrs);
3343 	KS_UINIT(txpkt_wrs);
3344 	KS_UINIT(txpkts_wrs);
3345 	KS_UINIT(txpkts_pkts);
3346 	KS_UINIT(txb_used);
3347 	KS_UINIT(hdl_used);
3348 	KS_UINIT(txb_full);
3349 	KS_UINIT(dma_hdl_failed);
3350 	KS_UINIT(dma_map_failed);
3351 	KS_UINIT(qfull);
3352 	KS_UINIT(qflush);
3353 	KS_UINIT(pullup_early);
3354 	KS_UINIT(pullup_late);
3355 	KS_UINIT(pullup_failed);
3356 
3357 	ksp->ks_update = update_txq_kstats;
3358 	ksp->ks_private = (void *)txq;
3359 	kstat_install(ksp);
3360 
3361 	return (ksp);
3362 }
3363 
3364 static int
3365 update_txq_kstats(kstat_t *ksp, int rw)
3366 {
3367 	struct txq_kstats *kstatp = (struct txq_kstats *)ksp->ks_data;
3368 	struct sge_txq *txq = ksp->ks_private;
3369 
3370 	if (rw == KSTAT_WRITE)
3371 		return (0);
3372 
3373 	KS_U_FROM(txcsum, txq);
3374 	KS_U_FROM(tso_wrs, txq);
3375 	KS_U_FROM(imm_wrs, txq);
3376 	KS_U_FROM(sgl_wrs, txq);
3377 	KS_U_FROM(txpkt_wrs, txq);
3378 	KS_U_FROM(txpkts_wrs, txq);
3379 	KS_U_FROM(txpkts_pkts, txq);
3380 	KS_U_FROM(txb_used, txq);
3381 	KS_U_FROM(hdl_used, txq);
3382 	KS_U_FROM(txb_full, txq);
3383 	KS_U_FROM(dma_hdl_failed, txq);
3384 	KS_U_FROM(dma_map_failed, txq);
3385 	KS_U_FROM(qfull, txq);
3386 	KS_U_FROM(qflush, txq);
3387 	KS_U_FROM(pullup_early, txq);
3388 	KS_U_FROM(pullup_late, txq);
3389 	KS_U_FROM(pullup_failed, txq);
3390 
3391 	return (0);
3392 }
3393